There’s a lot of attention on how AI can help model our genomes, proteins, and cells. But much of what shapes real-world health outcomes happens outside of our DNA, and rather within the trillions of microbes living in our bodies.

Today, that layer remains one of the most complex and most under-modeled systems in human biology.

That’s what Outpost Bio is tackling: building frontier models at this interaction layer.

Today, we’re thrilled to announce that Merantix Capital has co-led Outpost’s $3.5M pre-seed alongside Seedcamp, with additional participation from OpenSeed VC, Defined, and strategic family offices and angels.

Outpost is led by a deeply impressive founding team: CEO Dr. Jenny Yang (Oxford PhD and former Marie Curie Fellow with a background in clinical machine learning), and COO Alex Merwin, former Head of Growth for Health & Bio Startups at AWS.

As Jenny put it, “Building the most comprehensive dataset in human microbiology gives us a true foundation layer for modeling how microbial communities interact with drugs and other chemical inputs.”

I recently sat down with Jenny and Alex to discuss why microbiology is the next intelligence layer in medicine, what a true moat looks like in techbio, and how Outpost can reshape drug development and consumer health over the next decade.

Our edited conversation is below.

Adrian: Most biological frontier models focus on human DNA. You’re building at the interaction layer across microbial systems. What does that mean in practice, and why are you focusing on this space?

Jenny: Most biological AI models focus on human DNA — our genomes, proteins, and cells. We focus on the trillions of microbes that live in and on us, which directly interact with chemicals and shape how we metabolize drugs, absorb nutrients, and respond to products. These microbes contain far more genetic diversity than the human genome and can vary dramatically between individuals, making them a major — and highly personal — driver of health. We’re building models that capture this interaction layer so we can predict how this hidden ecosystem shapes real-world outcomes.

Adrian: How does Outpost Bio move beyond traditional microbiome research approaches?

Jenny: Traditional microbiome research has largely been limited to correlations, identifying associations between microbes and health outcomes, often relying on low-throughput systems like mouse models or bioreactors that can’t capture the diversity of real human microbial communities. Our lab-in-the-loop platform instead tests diverse, human-derived microbiomes against defined chemical perturbations at scale, directly measuring how drugs and compounds are transformed across different ecosystems. Machine learning then analyzes this highly dimensional, combinatorial space to pinpoint causal microbial drivers and determine the most informative next experiments. This closed feedback loop moves the field from correlation to mechanism, enabling us to model human microbiology at a scale and precision that wasn’t previously possible.

Adrian: Many techbio AI companies talk about proprietary data sets. What is the true moat here?

Alex: In practice, data alone is rarely defensible — it depreciates, it can be replicated, and in microbiome specifically, public datasets are growing fast. Our moat isn’t a static dataset. It’s the closed loop between wet lab and ML that generates new data designed to be maximally informative for model training.

We run high-throughput perturbation experiments on diverse, human-derived microbial communities — stool-derived communities, or SDCs — that capture the real complexity of the gut. Each experimental cycle feeds our models, and those models tell the lab exactly what to test next to fill the gaps in our understanding. That feedback loop means our dataset compounds in value with every iteration, and it’s tightly coupled to the models trained on it. You can’t just download it and replicate the insight.

On top of that, we’re building the infrastructure to make this data ML-ready at scale — enforced metadata, standardized representations, and a benchmarking suite that we intend to open-source. The companies that define how the field measures progress tend to stay at the front of it.

Adrian: Where do you go first as your commercial wedge and why?

Alex: Every oral drug is exposed to the gut microbiome before reaching systemic circulation. Microbial enzymes can activate, inactivate, or transform compounds — sometimes generating toxic metabolites or contributing to inter-patient variability in response.

Despite growing evidence, microbiome–drug interactions are not systematically profiled in preclinical development. There is no high-throughput, diversity-aware screen that captures how compounds behave across real human microbial communities.

We provide that missing layer. Our platform measures and models how compounds are metabolized across diverse microbiomes early in development — flagging potential toxicity signals, identifying variability risk, and informing formulation and clinical strategy before costly trials begin.

We’re starting with pharma because the unmet need is clear, the economic stakes are high, and structured pilot partnerships allow us to generate the high-quality data that continuously improves the platform.

Adrian: You describe building the most comprehensive dataset in human microbiology. What does that sort of data layer enable, and where do you see Outpost building from here?

Jenny: Building the most comprehensive dataset in human microbiology gives us a true foundation layer for modeling how microbial communities interact with drugs and other chemical inputs. Instead of isolated studies or correlations, we generate structured, high-throughput, causal data that allows AI models to learn transferable rules of microbial–chemical interaction. That data layer enables predictive tools for pharma, food, and consumer health companies to anticipate metabolism, toxicity, and variability before going into costly trials. From here, we see Outpost expanding into broader human microbiomes and becoming the intelligence layer that powers microbiology-driven product development.

Adrian: In 10 years, what looks different in consumer health or medicine because Outpost Bio exists?

Jenny: In 10 years, taking a drug or trying a new health product won’t feel like guesswork. Companies will be able to predict how different microbiomes metabolize a compound before it reaches patients or consumers, reducing unexpected side effects and variability in response. Personalized recommendations, from prescriptions to probiotics to skincare, will account for your microbial ecosystem, not just your genetics. Outpost will help make personalized medicine a reality by expanding it beyond human genomics to include the trillions of microbes that shape how our bodies actually respond.

Adrian: Thanks to you both. We’re excited to see what the future holds for Outpost Bio!

Check out Jenny’s recent panel at the Merantix AI Campus.

There are a growing number of very technical people — particularly at places like MATS, Apollo, DeepMind — who care a lot about AI going well. There are also a growing number of founders building with AI who want to build important things. Despite some examples like Goodfire (valued at $1.25B and backed by Menlo and Lightspeed), there’s still a coordination gap between the people who deeply understand AI risk and the people who know how to build venture-scale companies.

As VCs and entrepreneurs, we need to make more big bets on a safe AI future. I hope this piece is a small first step.

This is a post by Matilda Glynn-Henley, investor at Merantix Capital.

What is AI Safety?

AI safety is broadly the field concerned with preventing or mitigating the harmful consequences that may arise from the development and deployment of artificial intelligence systems. Within AI safety is the AI Alignment problem (making AI systems do what their creators intend them to do), ensuring the security and robustness of AI systems in deployment, and the more general risks that emerge as a result of AI systems proliferating in the real world. Think biosecurity, pandemics, AI-enabled cyber attacks, disinformation, and destabilization posed by a world where AI causes mass unemployment.

The landscape of opportunity here is wide. For example, as AI starts to displace white-collar jobs across Europe and the US, even figuring out how to reskill workers becomes an AI Safety problem. A safe AI future likely doesn’t include rapid and mass unemployment.

It is worth bearing in mind however that many of the hardest technical problems in AI safety — particularly those related to aligning future, more capable systems — may be better suited to researchers in research organizations or frontier labs than VC-backed startups.

Where should founders start?

One of the first steps at the inception of a startup is customer discovery. For founders hoping to build in AI Safety, one early consideration is whether you should try to sell solutions to the companies at the frontier of AI development, such as OpenAI or Anthropic, or not.

The frontier labs are not an easy or ideal customer for an early-stage startup. A good way to think about this is by considering their incentives. Why would a frontier lab outsource some aspect of their AI safety work to a third party?

At a recent Merantix Capital AI Breakfast panel discussion I hosted at the London AI Hub, we came up with some good reasons. Some AI infrastructure startups are being heavily funded on the expectation that frontier labs may prefer to acquire best-in-class teams later rather than build the capability in-house. Companies like Andon Labs, Goodfire and Apollo Research perhaps fall into this category.

You can also consider cases where regulatory pressure (e.g. around red-teaming or evals) or operational complexity makes it strategically cleaner to outsource certain functions to startups. For example large-scale Reinforcement Learning from Human Feedback (RLHF) operations (as Mercor’s trajectory illustrates).

The benefit of selling to frontier labs is that it keeps you “close to the action,” and AI capabilities at these companies are generally most advanced. They likely have a good indication of what the AI systems we deploy in the future will look like, and working directly with them gives an AI safety startup an unfair advantage in anticipating upcoming threats.

Nevertheless, having only a handful of potential customers is a high-risk strategy. Your commercial success becomes dependent on what labs decide is a priority.

Then where should AI Safety startups sell their solutions?

If your customers are not the big labs, what does your business look like? I see three broad paths (I’m sure there are more)

You could sell security, monitoring, governance, or guardrail software to companies deploying AI. This is what Ilia and Peter, two of our panelists at the AI Breakfast, were doing.

Lakera (acquired by Checkpoint for ~$300M) was also following this path, focusing on run-time security of LLMs, inference time protection, and agentic workflow monitoring. As AI gets deployed more widely, it would be a fair assumption that security for AI will become more of a topic and will probably require new startups to solve these new problems. AI is fundamentally non-deterministic and has a pretty wide and new attack surface. It is likely we’ll need companies here securing this technology. Problems like scalable run time monitorability outside LLM as a judge still remains unsolved.

If you are senior on an AGI team at a frontier lab, you could start another frontier lab designed to build AGI safely (think Safe Superintelligence SSI). Of course there’s a question of whether that mission drifts, as some believe was the case with Anthropic (The Atlantic, Jan 2026).

Let’s say that you are not a senior researcher at a frontier lab. You could focus on a vertical you know well and use safe and secure AI as a unique selling point where applicable. For example: secure voice agents for regulated industries (Gradient Labs’ focus on financial services), advancing the general knowledge of how to deploy AI safety in specific production environments.

Another way to think about AI safety is to consider the broader set of problems that arise in a world where AI is everywhere, and launch a company based on those assumptions. A few examples:

• Biosecurity and pandemic prevention: AI likely increases the risk surface for pandemics, so we’ll need technologies to prevent or contain them. Valthos, for example, is developing frontier AI systems that identify biological threats and design medical countermeasures in real time.

• AI for cyber defence: AI enables an uplift in capabilities of cyber criminals. Market trends already imply this, particularly as models lower the barrier to entry for sophisticated attacks and AI-generated code is massively prone to vulnerabilities (see White Circle).

• Insurance for AI: Mitigating the economic downsides of deployed AI systems going wrong, or building standards for secure AI (e.g. AIUC raised $15M) so that failures are less likely and easier to price in the first place.

• Diffuse and democratize intelligence: Workshop Labs, for example, builds individualized models (”one for each person”) to address what they call the intelligence curse aiming to distribute the benefits and control of AI more broadly.

• Defence of critical infrastructure: Cyber attacks on systems like the electrical grid are extremely high-impact, and we will likely need AI-native defences as the threat surface increases.

The Venture Scale Dilemma

Entrepreneurship, in particular venture backed entrepreneurship, is at its most basic level about finding big and important problems to solve. For VC investors, the problems must be large enough (so called “venture scale”) and customers must be willing to pay for them in order to justify a large return on investment. Not just a large return, but a venture scale return in the billions.

Venture-backed startups are absolutely not a solution to all problems in today’s AI safety landscape, but strong commercial incentives and the ability to scale makes them a good medium for some of these challenges.

Final Thoughts

I think there is a real opportunity to fund and found in AI safety, and I’m excited to see the next generation of AI safety founders working to solve these problems.

I also want to point to a few resources on AI Safety that I have found helpful.

If you are new to AI safety:

• Blue Dot Impact (AGI Strategy course is a good place to start, or Technical AI Safety for engineers and researchers)

• Apart Research Hackathons - if you are technical.

• AI Safety Founders - find co-founders in AI Safety

For deeper landscape views and other perspectives:

• AI Alignment & Trust Report

• AI Safety as a YC Startup

• Apart Research - where we are on for profit safety

• AI Safety and Entrepreneurship

If you are interested in building in the AI Safety space, or have your own ideas that I didn’t identify, please reach out to me at matilda @ merantix . com

We just returned from an action-packed week of discussion with AI leaders across industry, investing, and research at AI House Davos.

We couldn’t help but notice that the AI narrative has shifted a bit in Davos since we first co-founded AI House three years ago. In our first year, AI was a shiny and mysterious “new” object. Last year, it was about trying to separate the hype from the reality. This year, we’ve begun calling it the “year of AI ROI”. We heard a lot of variations of, “What can AI do for my business?” A real push for tangible, bottom line results.

That’s just one of our themes for the year. Below are 9 more. Ultimately, Merantix Capital is an investor. Our portfolio = our predictions. Nevertheless, here are a few educated guesses of what we think will shape AI in 2026.

For this post, I asked for some help from my colleague Thomas Wollmann, the CTO of Merantix Momentum, our AI solutions business.

1. Sovereignty and the European AI stack get real

The debate around European AI sovereignty is finally getting some momentum. One big topic in Davos was European Commission president Ursula von der Leyen officially endorsing the EU Inc initiative, an initiative we signed at its 2024 launch, which would institute the same capital regime across Europe and make it much easier for startups to register here, raise capital, and incentivize talent.

This should be the bare minimum, but it’s a real step in the right direction.

Meanwhile, by the end of 2026, we expect multiple European foundation models to be essentially good enough for most enterprise use cases, and for there to be a greater emphasis on European sovereign cloud providers. In that regard, we expect the narrative focus to shift away from whether Europe can (or should) build its own models, and more towards how the ecosystem actually works together, with a greater emphasis on infrastructure, deployment, public-private partnerships, and distribution.

2. Beyond (Transformer) LLMs

Three years after ChatGPT took the world by storm, we think 2026 will finally be the year where the conversation catches up with the research world and pushes past LLMs, particularly transformer LLMs.

LLMs are powerful interfaces, but their cost and weak grounding in the physical world are pushing researchers and companies to chip away and replace parts of the stack. LLMs will not go away, but state space models and flow models will play a bigger role, moving beyond transformer architecture.

And when it comes to the layperson’s understanding of AI, more people will see AI deeply implemented into their workplaces and personal lives in other modalities, shifting how they understand and talk about what AI really is

3. The Year of AI ROI

As we saw in the marketing installations along the Davos promenade, businesses are pushing to convey how AI can help the bottom line in real terms.

Enterprises are done asking if they should use AI. Now they want to know what it returns. We expect this year that more companies will be able to quantify this, measuring AI’s impact on the P&L. Earnings reports won’t just mention AI, they will quantify productivity as AI implementation becomes a board-level priority. Companies that cannot demonstrate AI ROI will stall.

4. Cybersecurity has a wake up call moment.

If 2025 was the year of “vibe coding,” 2026 is shaping up to be the year of security breaches. As AI-generated code and autonomous agents go mainstream, attackers are gaining leverage. We expect to see more scenarios where one skilled hacker (armed with AI) takes on an entire security team. Will 2026 produce “the big one?” A major AI-enabled cyberattack that forces enterprises and governments to completely rethink their assumptions around software trust, security, and resilience? We think so.

Every threat is an opportunity, which is why we back entrepreneurs with deep domain expertise in cybersecurity working to address these challenges head on.

5. AI’s job impact becomes impossible to ignore

For years, people have talked about the labor impact of AI in theory. So far, it’s been largely marginal or role-specific. Last year you started to see more storylines about how junior or entry-level white-collar work has been hollowed out, especially in knowledge work like software, consulting, or marketing.

This year, we think things will become significantly more noticeable. New junior roles in knowledge work will disappear as AI juices productivity gains in functions like customer support, sales, and content. For high-growth startups, this will mean scaling revenue without proportional hiring. For traditional businesses, it will mean layoffs or headcount freezing.

As a result, the importance of upskilling and reskilling and the future of education will become a more clear topic of discussion, as both new and existing workers in the job force look for alternative roles.

6. World models

If “beyond LLMs” is one theme, the natural follow up question is: What next?

Much of the talk at AI House, including a session with Yann LeCun, centered on embodied AI and world models. World models learn how objects move and how actions lead to consequences, making them foundational for general-purpose robotics. At the same time, world models can also power digital twins for factories, supply chains, cities, and financial systems, enabling decision making through simulation.

The next big leap will not come from words, but from machines that can imagine, simulate, and reason about the world before acting in it.

7. Death of search and and an e-commerce revolution

Consumers are increasingly asking, deciding, and buying through AI.

This may spell the slow death of the open web. But it also unlocks entirely new business models: LLM-native search, agentic commerce, and AI-first marketplaces that are not optimized for clicks.

2026 will be an inflection point. As consumer habits tip toward AI-driven shopping, the economics of the internet and who captures the value will get rewritten.

8. Constraints will drive the next wave of innovation

Constraints breed creativity. The era of “just add more GPUs” will have to end at some point.

We expect that constraints, whether they are geopolitical or financial, will turn compute into an even bigger battlefield in AI with more than just one dominant chip vendor winning via brute-force scaling.

Meanwhile a wave of innovation is expected to hit the hardware side this year. New, specialized chips will challenge the incumbents. On the software side, advances in inference optimization (compression) and efficient model design will push what’s possible with limited compute.

Doing more with less will be the name of the game, which will reshape the power dynamics and dependencies within the AI stack

.

9. AI for Science breaks through

AI is a powerful scientific assistant. That’s really just the beginning. Rather than just helping analyze or generate, AI systems are starting to decide themselves what experiments to run next, creating hypotheses, simulating outcomes, and iterating autonomously. The upside here is huge: AI systems that produce new knowledge and self-improve to accelerate progress.

This will have major implications in materials science, physics, mathematics, and beyond. We anticipate at least one major scientific breakthrough this year where AI is not a footnote or an assistant, but the primary driver. This will get significant press attention, accolades, and help people understand AI’s potential in the scientific realm.

10. And the Oscar goes to AI…

AI-generated content is having an interesting moment. On the one hand, our feeds are full of “AI slop” meant to farm engagement (like videos of Friedrich Merz doing a TikTok dance). Still, some AI content is indistinguishable from human output.

Oftentimes the dam breaks in the commercial realm. Coca-Cola released holiday ads created with the help of AI, which earned the soda maker some public criticism, but was more cost efficient.

We think 2026 is the year society will start to accept that (some) AI content can be high-quality. AI interviews will appear in major newspapers. AI-assisted research will distinguish itself in conferences. AI art will break through. An AI-generated film will receive awards.

This has happened with the advent of nearly every new technology. Radio scoffed at television. Print laughed at the internet. Once the quality bar is met, the acceptance of AI generation in the entertainment realm will unlock new formats and personalization.

Looking Ahead

Once again we couldn’t be more excited to see what this year brings in AI. What are we missing? And if you’re building something that helps solve the world’s most pressing issues with AI, let us know.

I’m very pleased to announce that FICUS Health has closed €3 million in funding, led by Redstone and following Merantix Capital’s initial investment through our venture studio.

Read more about the news in Handelsblatt.

When Benjamin Pochhammer first joined us on the AI Campus, he was looking into the intersection of AI and health, having previously founded CASPAR Health. Very quickly, Benny demonstrated not only passion for improving the lives of practitioners and patients, but also that he has the deep knowledge and expertise to build something big in this crucial area.

FICUS Health was born, centering on the rehabilitation space. Why? As Benny describes it, “Rehabilitation is a highly specialized environment with multidisciplinary teams, strict documentation rules, and a unique payor landscape dominated by the German Pension Fund. You can’t succeed here with a generic AI solution. You need deep domain knowledge.”

It’s a critical time, with clinics facing rising costs and workforce shortages. That’s why FICUS is so essential for practitioners, giving them more time to care for patients and reducing documentation effort by 70%.

FICUS has been off and running, working with nearly 100 rehabilitation clinics across Germany already.

I recently sat down with Benny and his cofounder Mario Elstner for a conversation about FICUS, how AI is impacting the patient journey, and where things go from here.

Our edited conversation is below.

Adrian: What is FICUS and why is it so important right now?

Benjamin: FICUS is an AI platform that transforms documentation processes in the rehabilitation sector. We automate one of the most time-intensive tasks in the clinical workflow, giving healthcare professionals more time for what matters most: their patients. For rehabilitation centers, FICUS becomes an enabler, helping them create efficient processes, maintain quality, and stay operational despite growing resource constraints.

The need has never been greater. The healthcare sector is facing a historic workforce shortage. At the same time, rising costs and tight budgets put additional pressure on medical centers while documentation workloads remain disproportionately high. And we haven’t even mentioned demographic change yet. Europe will see a massive increase in people over 65, driving chronic conditions and dramatically increasing the need for rehabilitation services.

FICUS directly addresses these structural pressures by giving time back to medical teams and enabling the facilities to operate more sustainably.

Adrian: How do you see AI impacting the health care space?

Mario: AI already impacts healthcare immensely, and the most immediate change is happening in the administrative space. Clinicians spend a disproportionate amount of time on tasks that don’t require their medical expertise. So, AI won’t replace healthcare professionals, but it will become a powerful agent that takes over repetitive work. Rehabilitation centers using AI will set new standards for efficiency and quality that others will ultimately follow.

Beyond that, AI will help connect the entire patient journey. Healthcare is still extremely fragmented; data and information gets lost between roles, systems, and stages of care. AI has the potential to turn these fragmented workflows into intelligent, adaptive systems that are interoperable by design.

This shift is one of the very few scalable solutions we have for addressing workforce shortages and structural constraints in health systems.

We’ve loved having the FICUS team on the AI Campus and as a really important part of our ecosystem.

Benjamin: From day one we have actively used the Merantix ecosystem. In the early stages it gave us a space to brainstorm openly and challenge our assumptions. And being part of Merantix also makes it much easier to attract top talent.

Mario: One moment that captures this collaboration was in the very early days. Before FICUS even had a real product, the Merantix team helped us shape and test our first MVP. This early support enabled us to enter the rehab market quickly, validate the concept with real users and build from day one with direct user feedback.

And we still benefit hugely from the broader community, from the many HealthTech events and panel talks to deep-dive sessions with other tech experts, founders and engineers. It is an ecosystem that keeps opening doors and brings the right people into the conversation.

What will you use the new funding for?

Benjamin: We’ll use the new funding to significantly advance our product and develop additional applications along the entire rehabilitation journey. A major focus will be interoperability, making sure FICUS connects seamlessly with existing clinical systems and becomes part of the daily workflow. At the same time, we’re strengthening our organizational backbone and scaling our commercial and implementation teams to expand our market position across the German-speaking region.

There’s a lot of competition in the AI health space. How are you taking on that challenge to differentiate and build a moat?

Benjamin: Our biggest differentiation is our focus. Rehabilitation is a highly specialized environment with multidisciplinary teams, strict documentation rules, and a unique payor landscape dominated by the German Pension Fund. You can’t succeed here with a generic AI solution. You need deep domain knowledge. Having built a successful tech company in the rehabilitation sector before with CASPAR Health, I’ve seen firsthand how the system works, where the real bottlenecks are, and what rehabilitation centers actually need.

Mario: We also build a moat through context and data. FICUS structures patient conversations, medical records, and anamnesis data in a way that gives our AI a uniquely detailed understanding of rehabilitation workflows. Combined with auditor insights and continuous feedback from clinicians, this creates a compounding learning loop that makes FICUS increasingly accurate and tailored to this space.

What do you think the health care space looks like in 5-10 years after AI is more fully implemented across health systems?

Mario: I think AI will fundamentally change the health care space over the next decades. It will support data collection about patients during anamnesis, augment their judgement in diagnosis drawing on knowledge bases and medical literature, and support with treatment suggestions and therapy monitoring leveraging historical data.

For patients, this has the potential to result in better access to medical information and care, more timely treatment and a greater focus on preventing medical issues in the first place. Right now we’re only taking the first small steps, but in the long run, for a sector that’s notoriously underfunded and under demographic pressure, AI may prove to be truly transformational.

Thanks very much, and congrats!

AI has become both cybersecurity’s biggest threat and its best defense. Attackers now use it to automate phishing, generate malicious code, and create deepfakes that can fool even advanced systems. Defenders are adopting the same technology to predict breaches, trace anomalies, and verify authenticity, but offensive methods still have a decisive advantage.

This tension defines the next wave of opportunity. Every advance in generative models expands the attack surface but also increases the need for smarter defenses. At the same time, it opens the door for new ventures building predictive, adaptive, and trusted systems.

In this article, I’ll explore where those opportunities lie and how founders can turn AI’s volatility into an advantage. I’ll also share insights from the recent “Deepfakes, Scams and Cybercrime” panel at the Merantix Campus in Berlin, where engineers, researchers, and founders discussed how these threats are evolving in real time, along with what that means for the companies racing to stop them.

AI’s Opportunity for Startups

AI is changing cybersecurity from a reactive discipline into a predictive one. It enables systems that model likely attack paths, anticipate exploits, and block them before they happen. It compresses the skills gap by embedding expert capability into software and brings enterprise-grade protection within reach of smaller firms. It is also unlocking new frontiers, from content authentication and fraud detection to the protection of autonomous systems. Despite the start of this defensive transformation, the scale of the challenge is enormous.

A World Economic Forum report found that 42 percent of organizations were successfully targeted by social engineering attacks in 2024. Nearly half of AI-generated code contains vulnerabilities, exposing products before they even launch. Model-serving APIs and large language model endpoints have become key targets, yet security tools for them remain rudimentary. Deepfake detectors perform well in controlled environments but fail in the wild. Together, these signals point to the same truth: attack surfaces are expanding faster than legacy vendors can adapt.

Even the titans of tech are feeling the squeeze. During the discussion at the Deepfakes and Cybercrime event, a former YouTube security lead in the audience described the pace of this dynamic: any detection system could be bypassed within six weeks. Their team measured success not by perfection but by halving damage with every iteration. That cadence should be a design goal for founders: ship, measure, retrain, repeat.

For founders, this widening gap represents both a technical and commercial opportunity. Security architecture built for static networks cannot protect self-learning systems. The next generation of cybersecurity companies will build for constant motion, predicting, learning, and neutralizing threats as they emerge. In that shift lies the foundation for an entirely new class of defensible: AI-native security ventures.

Where Founders Can Build

Each layer of the tech stack, from prediction and prevention to authentication and autonomy, is being rebuilt for a world where threats evolve faster than any human can track. The founders moving first in this shift are already defining how the next generation of security will work.

Predictive defense sits at the center of this change. Most existing tools still react after a breach, generating alerts once the damage is done. AI enables a different approach: systems that anticipate exploits, simulate attacks, and disrupt them in advance. BforeAI, based in France, is an example. Its predictive threat intelligence platform analyses billions of data points to identify malicious domains before they go live, serving clients that include Fortune 500 firms and NATO.

Revel8, from our own portfolio, uses AI to quantify exposure and simulate potential financial impact. The goal here is time compression. If attackers iterate in weeks, platforms need detection, simulation, and policy updates measured in days. Julius Muth, Co-Founder of Revel8, pointed to recent call-center breaches at large European insurers as examples of how scams now exploit workflows as much as systems, which makes predictive controls more valuable than reactive alerts. But not all companies enjoy the luxury of a robust security team.

Smaller companies are disproportionately exposed to cyber risks: one in three was attacked last year, with average breach costs of $255,000. Nonetheless, fewer than a third employ dedicated security staff. Larger companies typically employ a Chief Information Security Officer (CISO), along with dedicated security teams, but this decades-old security gap between SMEs and enterprises could soon come to an end. AI now makes it possible to embed those functions into SaaS platforms.

Cynomi has harnessed this model with an AI-CISO platform that automates assessments, generates policies, and provides compliance dashboards. In 2024, it signed deals with more than 100 service providers, including Deutsche Telekom, distributing its platform to thousands of SMBs, showing that these services serve a real need and a huge addressable market. These platforms should embed opinionated process controls, guardrails that automatically enforce secure behaviour inside workflows, such as requiring multi-party approval for payments or verifying caller identity through a trusted channel.

Trust infrastructure represents another frontier. Cheap deepfakes and even “deepfake-as-a-service” make detection alone a losing strategy, something that Hasna Najmi, Senior AI Solutions Architect at Merantix Momentum, pointed out during our panel. Najmi said that these services are so efficient that turnkey providers can now onboard users via chat and generate realistic videos in minutes. Detection also remains probabilistic. Muth explained that CISOs must still define thresholds and decision rules for false positives and negatives, while Dr. Nicolas Müller, Senior Engineer at the Fraunhofer Institute, noted that audio deepfake detectors fail to generalize to new model families. Müller added that if generative AI consolidates around a few providers, watermarking and training on real outputs could become practical, improving provenance and verification at scale.

Bigger threats, bigger opportunities

AI is both the nemesis and knight in shining armour for cybersecurity. But it’s also more complicated than that: AI adoption itself is also creating risks deeper in the infrastructure layer. Coding assistants now write large portions of production code, often with hidden vulnerabilities that pass untested into live systems. APIs are multiplying, model-serving endpoints are exposed, and the infrastructure that supports AI models has become a new target. The OWASP 2025 LLM Top 10 and NIST’s Generative AI Profile both highlight the same weak points: injection flaws, insecure outputs, and supply-chain risks linked to AI-generated code.

The consequences are already visible. Ray, an open-source framework designed to scale AI and Python applications, experienced a hack in which its clusters were hijacked for crypto-mining in 2024, and NVIDIA’s Triton inference server required an emergency patch in 2025 after attackers found remote-execution flaws. These problems continue to exist at scale, but a few European startups are beginning to respond. HarfangLab has grown quickly with its AI-powered endpoint detection platform, now surpassing €40 million in annual revenue, while Lupovis, spun out of the University of Strathclyde, uses deception-based AI honeypots to lure and trap attackers.

For founders, the opportunity runs deeper than patches or point solutions. The market still lacks foundational tools that secure AI code and infrastructure from the inside out: firewalls, continuous validation systems, and automated runtime protection built for a world where models write and deploy themselves.

Autonomy and counter-drone security are becoming the new front line of cybersecurity. As the boundary between digital and physical systems dissolves, attacks on drones, satellites, and autonomous vehicles can now trigger real-world consequences. The war in Ukraine exposed how quickly off-the-shelf drones, powered by commercial AI models, can be repurposed into weapons or surveillance tools. The UK NCSC has warned that AI will shorten exploit timelines for critical infrastructure, while NATO and EASA have elevated counter-drone initiatives as strategic priorities.

This shift turns autonomy itself into an attack surface. AI systems that once operated in isolation—navigation, perception, targeting—are increasingly networked and therefore vulnerable to interference, spoofing, and data poisoning. Keeping these systems secure demands continuous model updates, telemetry, and field learning. Europe is beginning to respond. Helsing, Quantum Systems, Dedrone and Origin Robotics represent the first generation of companies building or defending AI-powered autonomy. Their rise signals both the urgency of the challenge and the gap still left to fill. For founders, the opportunity lies in securing this frontier: embedding cybersecurity into the fabric of autonomous systems before adversaries do.

Why This Matters for Europe

Europe’s growing focus on digital sovereignty is reshaping the cybersecurity market. The AI Act sets new standards for trustworthy and transparent AI, while the NIS2 Directive raises security requirements for critical sectors. Together, they are tightening compliance expectations across the continent and creating strong demand for trusted infrastructure. Nearly half of European organizations plan to adopt sovereign cloud solutions by 2025, yet the market remains dominated by a few non-European providers. This imbalance leaves governments and enterprises exposed to both cyber threats and external jurisdiction through laws such as the U.S. Cloud Act.

Momentum is now shifting toward self-reliance. Initiatives like Gaia-X and proposals for a European Tech Sovereignty Fund reflect how sovereignty has become both an economic priority and an investment theme. Enterprises and critical-infrastructure operators are seeking EU-governed systems that guarantee control over data and operations. For founders, this environment provides a natural tailwind. Every new rule, funding program, and procurement policy strengthens the case for European-built, AI-native security platforms.

The same dynamic is playing out among smaller companies. Regulation is pushing for local control, while AI is making it possible to deliver advanced security capabilities at scale. Automated platforms can now offer CISO-level oversight at a fraction of the cost, giving SMEs access to protection once reserved for large enterprises. As these tools spread through supply chains and business networks, they will define how Europe achieves digital autonomy and long-term resilience.

Build the future of cybersecurity with us

The cybersecurity arms race is already underway, and the winners will not only protect Europe but also define new global categories. The challenge is immense, but so is the upside. Founders who build sovereign, AI-native platforms that predict and neutralize threats, democratize security leadership, and anchor trust infrastructure will create the next generation of cybersecurity leaders.

In our venture studio, we partner with founders (like revel8) who are ready to take on this challenge. If you are thinking about building in this space and want to shape the future of AI-first security, get in touch with us.

Vara, a Berlin-based health tech company specializing in breast imaging, has received the only CE certificate on the market for an AI that can serve as an independent second reader in mammography, both for screening and diagnostic use.

Across Europe, national screening programs face a capacity crisis: more women are eligible for breast cancer screening than ever before, while the number of radiologists declines.

The “double-reading” standard, where two radiologists independently interpreting each mammogram, remains best practice. But as workloads rise by 30–40%, this gold standard is at risk of becoming unsustainable.

AI as an independent second reader has been expected for years. Until now, no system had met the regulatory, ethical, and evidence bar to make it a reality.

How Vara got there step by step

Independent AI in healthcare isn’t mainly about writing code. It’s about earning trust.

Vara’s path reflects a decade of deliberate progress through the four milestones that regulators and clinicians have long demanded:

a) Data foundation: Vara partnered with screening centers across Germany, Europe and emerging markets to build one of the world’s largest curated breast-imaging datasets with far more than 15 million biopsy- and 2 year follow-up proven mammograms.

b) Retrospective validation: Early models showed strong performance but, like most AI in medicine, still lived in the “retrospective” world – safe in research, hard to translate into practice.

c) Clinical integration: Vara then embedded its AI as a support system (“concurrent read”) across the German National Screening Program, now covering ~40 % of all centers. This workflow allowed radiologists to use AI safely while building real-world evidence across a whole nation.

d) Real-world prospective evidence: In 2025, Vara initiated and co-authored the PRAIM study, published in Nature Medicine – the largest real-world, prospective AI study in healthcare. Across 460,000 women screened in 12 centers, PRAIM demonstrated statistically significant improvement in cancer-detection rate (+17.6% in cancer detection rate), reduction in false positives (-20%), and meaningful workload reduction for radiologists (-56.7%).

e) Building real-time monitoring: If an AI system takes independent decisions, it is crucial (also as part of the new EU AI Act) to ensure human oversight and understand how real-world adaptations impact the models’ performance. Vara is the only company on the market with a real-time monitoring system built into its software.

Only after that sequence – data, validation, integration, real-world evidence, and accountability – did Vara apply for regulatory recognition of AI as an independent read.

What’s behind the CE certification

The new CE mark, issued under Europe’s Medical Device Regulation (MDR), authorizes Vara’s AI to act as an independent second reader in mammography.

That means:

• In a double-reading program, one of the two reads can now be performed autonomously by AI.

• In diagnostic workflows, where a second human read is often no longer economically feasible, AI can now restore that safety net.

This is not “decision support.” It is an independent diagnostic interpretation, regulated as a high-risk Class IIb medical device under MDR.

Equally important, the certification acknowledges Vara’s real-time performance-monitoring system, which tracks how the AI behaves across populations, scanners, and time.

That transparency is what makes autonomy possible.

What’s next

Regulatory approval is the beginning, not the end.

For AI to play its full role in screening programs, national guidelines and reimbursement frameworks must evolve.

In Germany alone, 3,000 additional cancers would be found early enough every year. The survival prognosis for early-stage breast cancer is very high (>90%) but drops significantly if found too late. Additionally, 30,000 unnecessary recalls (with enormous emotional pain associated with them) could be prevented.

Germany is uniquely positioned to lead.

Its screening program has just completed a 10-year evaluation confirming significant mortality reduction. Its health-technology-assessment (HTA) framework provides a clear, evidence-based process for evaluating new methods.

The opportunity now is to adapt that process for AI.

If successful, Germany could become the first country in the world to formally integrate AI as a standalone reader in a national cancer-screening program.

A decade ago, AI in radiology was overhyped. Today, it’s finally proven.

And with this CE certification, the field moves from speculation to implementation – responsibly, transparently, and at scale.

The logistics industry is chaotic and full of manual planning, spreadsheets, and phone calls.

It’s the kind of terrain ripe for AI and automation, and a space we’ve been really interested in at Merantix Capital.

Which is why I’m thrilled to introduce our investment in Arqh, which today announced its $3.8 million pre-seed financing round led by our friends at Founderful.

Check out the story here.

Arqh fuses language models with a proprietary optimization engine, a unique agentic system that continuously re-plans real-world logistics operations and explains its choices to keep human dispatchers in the loop.

I recently sat down with Arqh’s founder Antonia Unger to discuss the current logistics landscape, how AI will reshape the supply chain, and where Arqh goes from here.

Antonia’s background in the logistics space showed her that there was a huge opportunity. “Working side‑by‑side with dispatchers, we saw firsthand what their day‑to‑day looks like and were shocked by the inefficiencies and dinosaur software being the status quo,” she told me.

Here’s our edited conversation below.

Adrian: What’s wrong with the logistics industry today?

Antonia: Critical planning — daily dispatch/orchestration and long‑term strategic decisions — remains largely manual. To automate and optimize these decisions, companies to date had two options: Off‑the‑shelf solvers that capture only a small fraction of real constraints (time windows, HOS, dock capacity, equipment compatibility, service levels, live vs. drop, cross‑border rules, etc.). Or bespoke systems that require 10+ MSc/PhD‑level specialists to build and continually retune as operations evolve.

The net effect is organizations revert to spreadsheets and phone trees. Not because they want to, but because current systems can’t represent the business as lived or re‑optimize fast enough when conditions change.

Our motto: Run your operations with zero spreadsheets and zero PhDs.

How does Arqh work? You talk about your system being both “right brain” and “left brain.” Can you explain what you mean by that?

To build the complete intelligence for logistics there are two requirements: handle the messiness of the real world and deliver mathematical precision. These are our two “brain halves.”

Our “Right Brain” is based on language models and understands policies, tribal knowledge, unstructured inputs and can capture intent and context. Our “Left Brain” is a rigorous optimization layer that encodes constraints, proves feasibility, and delivers measurable outcomes, i.e. the “logic & math half.”

Only when both work in tandem can you make complex optimization usable for non‑technical, rightfully skeptical industry veterans. And as dynamic as the real world, not a static one‑time algorithm.

Can you walk me through a very clear use case of Arqh in plain terms, if I’m not someone super knowledgeable about how logistics works?

One concrete use case is daily dispatch planning for a regional last‑mile fleet (e.g., northern Switzerland).

Without Arqh: At 17:00, a team of planners exports orders from the ERP and begins planning 8–15 trucks per person, based on orders and (optionally) service‑level agreements (SLAs). It’s a long, iterative process to satisfy resource, customer, product, and legal constraints. When execution starts at 07:00 the next day, an employee calls in sick, a dock runs late, traffic builds, or a customer doesn’t open. The plan breaks and firefighting begins.

With Arqh: Arqh ingests orders, driver hours‑of‑service (HOS), equipment, dock calendars, SLAs, and live feeds (traffic, weather, carrier updates). It quickly generates a baseline plan that adheres to all constraints and is 10%+ more efficient than a human‑generated plan.

As disruptions hit, the planner communicates in plain English and re‑optimizes the plan.

The result is faster planning, fewer manual calls, fewer empty miles, better on‑time performance, and calmer operations when things change mid-day.

Other use cases where Arqh supports decision‑making include: Strategic what‑if analysis of warehouse locations, fleet mix, and customer tariffs. Streamlined communication among managers, dispatchers, drivers, and end customers. And easy onboarding of new customers and operational changes.

What’s your background and how did you first identify this problem in the logistics space?

My co-founder Mert and I have an ETH Zurich / EPFL background. We both worked in consulting and organically ended up partnering on multiple logistics and transportation projects. Working side‑by‑side with dispatchers, we saw firsthand what their day‑to‑day looks like and were shocked by the inefficiencies and dinosaur software being the status quo. It makes sense as logistics hasn’t historically attracted as much builder attention, but the gap was bigger than we expected, so we started Arqh to close it.

What is the go to market plan for Arqh?

We’re taking a verticalized approach, focusing first on B2B last‑mile logistics. We’re seeking strategic partners in each sub‑vertical to co‑develop technical blueprints of the optimizer and ensure we cover all requirements in those sub‑verticals.

And market timing is perfect. Language model capabilities are exploding but they lack logic and haven’t yet been connected with powerful, custom optimization. We’re building that bridge.

I’m Swiss, so maybe a little biased here… but can you talk a bit about the Swiss AI ecosystem and how it’s matured and grown in order to help you all succeed?

Swiss talent is highly skilled, and universities are open to collaboration. From very early on we worked with research labs at EPFL and kept close ties with the ETH AI Center. Although ultimately customers are what count, the ecosystem is buzzing, the network is strong, and the talent hungry. We wouldn’t want to build Arqh anywhere else.

Where do you see the future of logistics going? What do these systems look like in 5-10 years?

Over the next decade, logistics will evolve into an AI-orchestrated, self-optimizing ecosystem. In 5–10 years, the work that today consumes armies of planners —dispatching, scheduling, capacity planning, rate discovery -—will be handled continuously by intelligent systems that learn and adapt in real time. Humans won’t disappear, but their focus will shift: they’ll set strategy, tune risk preferences, and manage relationships.

Because data standards and interoperability will have matured, these systems won’t just optimize within a single company, but across entire supply-chain ecosystems. The result are networks that are leaner, greener, and more resilient. Margins strengthen, emissions fall, and operators spend their time steering the system instead of fixing it.

What will you use the funding for?

Growing our core engineering team in order to be able to develop the language layer and optimization blueprints more rapidly in different sub-verticals.

Taking this verticalized approach will get is where we need to be to be able to really disrupt.

Thanks very much for the chat! We can’t wait to see where Arqh goes from here.

The earliest applications of AI have made us used to thinking of it in the context of software: intangible, abstract, algorithmic. Or how I just asked ChatGPT to make the picture above for this article.

But AI’s most transformative applications will most likely emerge in the physical world, where it can reshape how we design and scale new materials.

That’s why materials discovery is one of our biggest areas of interest at Merantix Capital. Our portfolio already includes Cambrium, a company born out of our venture studio that designs novel proteins.

If you are in Berlin, we are hosting an AI Breakfast on Thursday 21st August at the Merantix AI Campus to explore these ideas in more depth, with speakers from Cambrium, CuspAI, Dunia, and our own GP Adrian Locher.

If you’re already building in this space (or are keen to), we’d love to see you there!

Why we like the space

Materials have long been the hidden foundation of progress, underpinning advances in semiconductors, solar cells, batteries, and carbon capture. Now, AI is accelerating every stage of their development. It proposes novel compounds, predicts their properties, and links design directly to manufacturing constraints. This convergence opens the possibility of building entirely new materials platforms that are faster, more cost-effective, and more adaptable than ever before.

The next generation of materials giants will fuse deep technology with industrial relevance and, most importantly, help address some of the most urgent challenges of our time, like clean energy and sustainable production.

For years, computational chemistry served as a tool to simulate a few candidate molecules before testing them in the lab. The process still relied on human intuition to select what to simulate but, today, large-scale AI models don’t just evaluate known materials. They propose novel ones at unprecedented scale.

DeepMind’s GNoME project, for instance, trained a graph neural network to predict the stability of inorganic crystal structures. Researchers at Berkeley Lab for their part built A-Lab, an autonomous discovery platform combining machine learning, robotic synthesis, and automated measurement. In a 17-day trial, it completed 355 experiments and discovered 41 new inorganic compounds, all guided by active learning algorithms that improved in real time.

Designing for Scale

Big as these developments in R&D, they’re only going to help drive forward science if they can be applied at scale. That’s why a new cohort of startups are building their discovery platforms with production in mind from the beginning.

Cambrium is taking this scale-oriented approach into the world of protein materials. The company uses AI to design novel proteins that do not exist in nature and produces them through microbial fermentation, allowing for a more flexible and sustainable production process. From the start, its design process has been built around manufacturability, with a focus on properties like thermal stability, fermentation yield, and compatibility with existing bioproduction infrastructure.

This approach led to a fivefold increase in development speed and a twentyfold reduction in cost compared to traditional protein engineering workflows. One of Cambrium’s first products, NovaColl, is a sustainable collagen alternative that achieves a 95 percent lower carbon footprint and was developed in under two years.

In climate-focused materials, CuspAI is using generative AI to design bespoke porous compounds for applications like carbon capture and gas separation. The platform functions like a “search engine” for materials, generating and evaluating candidates that match specific performance criteria, including CO₂ selectivity and release dynamics. This approach has significantly shortened development timelines, reducing discovery cycles from years to months and achieving success rates of around 90 percent in generating viable materials. While still early, CuspAI is prioritising industrial deployment by designing materials that not only perform well in theory but also meet practical requirements such as regeneration efficiency and compatibility with existing capture systems.

Materials Nexus is another example addressing the supply chain risks associated with rare-earth elements by using AI to design alternative permanent magnets. In collaboration with the University of Sheffield and the Henry Royce Institute, the company screened more than 100 million candidate materials to develop MagNex, a rare-earth-free magnet suitable for use in electric vehicles and wind turbines. The material was taken from concept to synthesis in just three months and can be manufactured at around 20 percent of the cost of typical rare-earth magnets, with a 70 percent reduction in carbon emissions per kilogram produced.

XtalPi may have started in pharmaceutical discovery, but its AI-powered R&D platform is now proving adaptable to materials science, including solid-state research. The company offers tools tailored to structure prediction, experimental solid-state chemistry, and automated crystallography, collectively marketed under the “XtalGazer” and digital chemistry suite. Since 2024, XtalPi’s public filings confirm that its platform is being deployed for electrochemical and new energy material development, such as battery electrolytes and advanced ceramics, often in collaboration with partners like GCL Group and Fangda Carbon. These developments demonstrate how an AI-native discovery stack, once proven in one sector, can be repurposed to serve adjacent industries with minimal changes.

Together, these companies illustrate a broader shift: materials discovery is no longer just about predicting what might work in a lab. It’s about building entire pipelines that move from concept to manufacturable product with speed, precision, and commercial focus. By starting with real-world constraints such as cost, scale, and performance, this new wave of ventures is showing that AI-native materials platforms can make a meaningful impact far beyond the simulation stage. The next question is where new founders can still add value.

Where Founders can build

Founders entering the materials space face distinct challenges and opportunities. Initial traction tends to emerge in industries where performance matters more than cost. Markets such as cosmetics, fashion, and specialty coatings offer fast sales cycles and customers willing to pay for novelty, differentiation, or sustainability. These sectors provide early revenue, which can fund further technical development and create defensible IP.

Over time, the opportunity expands into heavier industries. Sectors like construction, packaging, agriculture and energy require greater investment in scale and compliance but offer outsized long-term value. In these markets, success depends not just on novel discovery but on the ability to integrate with supply chains and meet regulatory standards.

Strategically, founders can choose different entry points across the stack:

• Some will focus on software platforms, offering prediction tools to chemical companies or research institutions.

• Others will partner with OEMs, co-developing materials while relying on partners for scale-up.

• A third category will go full stack, owning discovery, validation, and production in-house.

Each approach carries trade-offs in capital requirements and speed, but all benefit from tighter control over learning cycles.

What is still missing in the ecosystem presents new possibilities. Lab automation remains capital-intensive and geographically limited. There is a shortage of multimodal models capable of integrating structural, textual, and experimental data. And very few infrastructure tools exist to bridge gram-scale discovery with ton-scale manufacturing.

In this landscape, speed of iteration becomes the key source of defensibility. Teams that can cycle faster, learn from data, and apply those insights across markets will build the next category-defining materials companies.

I’m always looking to hear from founders in materials discovery — if this article speaks to you or you have an idea you want to develop, get in touch or come along to our AI Breakfast this Thursday in Berlin.

Check out the great coverage in Handelsblatt, Tech EU, and Tech Funding News.

Here’s why we’re really excited. When you look at the AI agent landscape, things are moving quickly with desktop and web workflows. But mobile, the dominant user interface, is still untapped. We’re soon going to be looking at a world where you don’t need to switch between six apps to check flights, book a hotel, file your expenses, and update your calendar and colleagues. You’ll tell your agent to “book my trip to Berlin next week for the client meeting.”

That’s what Droidrun is building, and the company has already gone viral for its open-source efforts, with 900+ developers signing up for the framework in 24 hours, collecting 3,300+ GitHub stars and becoming the fastest-growing mobile agent project.

Droidrun does that by taking a different approach. “We give the AI the exact structure of what's on screen,” co-founder and CEO Christian Ninstel explained to me recently. “It's like creating an API for every mobile app that's ever been built.”

We’ve been wowed by Droidrun’s speedy traction and the execution abilities of the team. I recently sat down with Christian to hear about Droidrun’s growth and where things go from here. Below is our edited conversation.

Adrian: You built Droidrun after hitting a wall with existing mobile automation. What exactly was broken, and how did you solve it differently?

Christian: We needed AI agents to retrieve product insights from viral TikTok videos for our previous startup. Sounds simple, right? But every existing mobile automation solution failed spectacularly on real apps. They all break when apps update their UI or when you need to handle dynamic content.

The problem is that they try to "see" screens like humans do, which is slow and unreliable. We took the opposite approach: converting mobile interfaces into structured text that LLMs can process directly. Instead of trying to find a button by its visual appearance, we give the AI the exact structure of what's on screen. It's like creating an API for every mobile app that's ever been built.

900+ developers signed up in 24 hours after launch. That's incredible traction for a dev tool. What drove that viral moment?

We launched with a bold claim: "AI can now use your phone." But more importantly, we had a working demo that actually delivered on that promise. Developers had been frustrated by the same mobile automation problems we faced.

The virality came from community channels like Reddit, X, and open-source visibility. We hit 1,000 GitHub stars in under 48 hours because developers could immediately see the value and start testing it.

To maintain growth, we're doubling down on community, developer experience, shipping cloud runtimes, and integrating with major agent frameworks like LangChain, CrewAI, and AutoGen.

Can you walk me through some concrete use cases where Droidrun is already creating value?

We're not building apps, we're building the infrastructure layer that powers them. Think of us as the execution runtime that other companies build on top of.

QA Infrastructure: For example, QA Startups can integrate Droidrun to offer "AI-powered mobile testing" to their customers. Instead of manually writing test scripts, QA teams just describe what they want to test in natural language, and our agents execute across real devices. We're seeing testing companies add mobile AI capabilities to their platforms in weeks, not years. We also discovered that our cloud infrastructure is what a lot of companies can use to scale up and decrease costs.

RPA Platform Extensions: Major automation platforms could integrate Droidrun to extend their desktop workflows to mobile. Imagine a business process that starts in Salesforce, moves to email, then needs to update a mobile-only app - that final step was impossible before. Now RPA providers can offer end-to-end automation that includes mobile.

What's interesting is that we're often faster and more reliable than desktop automation. Desktop RPA deals with complex browser environments, bot detection systems, and constant UI changes. Mobile apps are more predictable; they have consistent UI patterns and fewer anti-automation measures. Our structured text approach bypasses the visual complexity that slows down desktop automation.

Agent Framework Infrastructure: LangChain, CrewAI, and AutoGen are integrating Droidrun so their agents can execute mobile actions. Developers building AI assistants can now say "and also handle mobile workflows" without building mobile automation from scratch.

Enterprise Automation APIs: Companies in fintech and logistics could use our APIs to automate mobile-only business processes. One financial services company built an automated compliance system that verifies transactions across 12 different banking apps - something that was previously 100% manual.

Developer Tool Platforms: Testing and monitoring platforms are white-labeling our technology to offer "mobile AI agents" as a feature. They focus on their core product, and we handle the complex mobile execution layer.

The pattern is clear: everyone wants mobile AI capabilities, but building reliable mobile automation is incredibly hard. We're the infrastructure layer that lets other companies ship mobile AI features without spending years solving the core execution problem.

Mobile is notoriously locked down, especially iOS. How are you handling the technical challenges of cross-platform support?

We already support iOS in private beta, but you're right - iOS is more restrictive. We had to work around stricter sandboxing and app store limitations.

Our approach is hybrid: Android for open experimentation and rapid iteration, iOS for enterprise integrations where we can work within corporate device management systems.

The key insight is that enterprises want this capability badly enough to work within platform constraints. Long-term, we're building toward a unified cross-platform runtime that's open, fast, and production-ready.

Where do you see the biggest opportunities for Droidrun in the next 2-3 years?

Short-term: High-friction industries where APIs don't exist or fall short - fintech, healthcare, logistics. These sectors run on mobile apps but lack automation infrastructure. Banks need to verify payouts, hospitals need to update patient records, logistics companies need to track shipments - all through mobile apps.

Long-term: We're building toward a world where mobile becomes an agentic operating system. Instead of opening ten apps to book a business trip, you express intent and agents execute across your entire mobile ecosystem.

The market timing is perfect - AI agents are exploding, mobile usage is dominant, but the two haven't connected yet. We're building that bridge.

You've raised €2.2M in a challenging funding environment. How are you planning to use this capital?

Three priorities:

1. Team expansion. Growing our core engineering team to ship faster.

2. Cloud infrastructure. Launching managed runtimes for Android and iOS so developers & enterprises can deploy at scale.

3. Ecosystem growth. Deep integrations with agent frameworks and developer tools, plus early proof-of-concept products.

The key is staying open-source at the core. That's what drove our initial traction, and it's how we'll build the ecosystem. The funding lets us accelerate without compromising on that developer-first approach.

Thanks very much for the chat, we’re excited to see where Droidrun goes from here!

What we’re starting to see now is that the real leverage is moving to the interface layer (as in: how people are actually interacting with AI). This is where users make decisions and generate the feedback that improves the system. It is where value is created and, for startups and investors, where it can be captured.

In this piece, I’ll be looking at why this shift is happening and how AI-native interfaces are becoming the new strategic layer across industries like productivity, legal tech, logistics, and fund operations.

The shift toward interface-level innovation is already happening.

OpenAI recently acquired former Apple designer Jony Ive’s hardware studio io for $6.5 billion. The real signal wasn’t the deal size, but the focus: ambient AI designed to blend into everyday life without screens. We’re witnessing interaction move from something active and controlled to something embedded and passive.

Another signal for the changing interface of AI was the ElevenLabs hackathon, where a team built GibberLink, a protocol for AI agents to communicate using compressed audio. It’s early, but it points to a world where agents coordinate directly, without waiting for human input. How will agent-agent communication be structured, especially on platforms in which humans typically interacted — for example, over the phone? What happens when your personal booking agent calls your doctor's office's booking agent to arrange an appointment? How they interact is a big question.

The shift is changing how people build. Tools like CustomGPT and Google AutoML are making natural language the default interface for software creation. The rise of “vibe coding”, describing an outcome in plain terms and letting the system handle the structure, is lowering the barrier to entry. Gartner estimates that by 2025, 70 percent of enterprise apps will be built using no-code or low-code tools.

Each of these examples offers a view into how interaction with AI is evolving. Interfaces are no longer just access points. They define how we unlock the value of backend AI systems, where a good user interface allows the true potential of AI functionality to be unlocked. The first steps have been made and this shift is already influencing how tools are being designed, adopted, and embedded into daily work.

Innovation and the decision layer

As AI moves from experimentation into real-world use, interfaces have become the main point of leverage. We can see this play out across three layers: general-purpose assistants, embedded enterprise copilots, and domain-specific platforms.

Chat-based assistants like ChatGPT and Claude’s strength lies not just in language generation, but in how they turn complex inputs into simple, guided exchanges. But that was just the beginning. Chatbots reduce user autonomy and increase cognitive load. Why should we ask AI to do something when ambient computing can unlock AI pre-empting actions based on what you are already doing.

OpenAI’s move with Jony Ive shows they are thinking about what comes next: how people interact with AI will (and needs to) shift to AI-native devices and processes.

Enterprise copilots take this further by embedding themselves inside the tools where people already work. Microsoft 365 Copilot and Notion AI are not trying to replace existing applications. Instead, they surface suggestions and automate tasks within the same interfaces teams already use to write, calculate, meet and plan.In other words, everyone is realizing that the interface is the control point. While one might think that reality favors incumbents, this also creates a huge startup opportunity. The challenge is that giants like Microsoft are more easily able to embed AI into daily workflows, but startups that are able to “own the interface” in underserved workflows, such as procurement, logistics, construction, R&D, etc., will be positioned to become new category leaders in every domain.

Indeed, in more specialized environments, vertical platforms are emerging that target specific roles and decision points. Libra, a legal tech and Merantix Capital portfolio company, supports contract analysis, drafting, and research in one system. This year, the firm reached over €1 million in ARR thanks to an interface that understands legal language, firm-specific workflows and decision logic.

Canoe is doing something similar in fund operations. It automates data ingestion from custodians and portfolio systems, flags anomalies, and creates compliance-ready reports. The system integrates directly into operational workflows, which makes it easier to trust and easier to adopt.

Interfaces are becoming decision environments. When that happens, they start to accumulate operational data, model edge cases and adapt in ways that horizontal platforms cannot easily replicate.

Vertical Interfaces: Where AI Meets Real Work

Some of the most important changes are happening in domains that have historically been underserved by software, from fund administration to supply chain planning and everything in between.

These workflows tend to be complex, manual, and spread across systems that were never designed to work together. That has created an opening for AI-native platforms that lead with interface design. I’ll give a few examples from outside of our portfolio.

Procurement is one of the clearest cases. Mercanis has built a system that combines sourcing, supplier management, and contract workflows into one environment. The platform automates routine tasks like RFx processes and risk checks and continuously analyzes spend data to suggest savings opportunities. According to EU Startups, the firm raised €17.3 million in June, and reported that some customers experienced a 40 percent process cost savings and a 2.5 times improvement in efficiency.

In supply chain operations, Pelico helps manufacturers and logistics teams detect bottlenecks, manage disruptions and coordinate responses across departments. The system integrates live data from enterprise resource planning (ERP) and factory systems, as well as proposing real-time adjustments. For companies like Airbus and Safran, this has led to a 40 percent reduction in missing parts and a 15 percent lift in on-time delivery. By giving teams a shared interface to make decisions and act quickly, the platform improves both visibility and execution.

Raft is doing something similar in freight forwarding. The platform replaces manual processes with an integrated workflow that handles shipment tracking, invoice reconciliation and customs documentation. A customer-facing portal provides visibility into bookings, emissions data and quotes. The firm recently raised $30 million in Series B funding and has gone on to become a daily workspace for logistics teams.

What these platforms have in common is not a particular technical breakthrough, but a focus on owning the interface where decisions happen.

By embedding themselves in the day-to-day work of operational teams, they gather data that improves the product, reinforce habits that are hard to unlearn, and position themselves as the primary system of record. In environments that are structured but nuanced, where decisions happen continuously, interface ownership will continue to be a durable strategic advantage.

Our thesis: Interface ownership will create new category leaders

Across every example, one idea keeps showing up: the most meaningful value accrues where decisions are made. Interfaces that support that moment, where information turns into action, end up shaping behavior.

That’s why AI-native UX consistently outperforms tools that simply layer models onto old workflows. Retrofitting intelligence into legacy software often misses the timing, the structure, or the real context of the task. In contrast, deeply integrated vertical copilots are proving more resilient by becoming part of the decision process. We believe that the next generation of category leaders will be built at this layer.

If you’re building interfaces that change how work happens, we want to hear from you.

For anyone building in AI, the Merantix AI Campus is a one-stop-shop for idea sharing, networking, and collaboration. So when I was introduced to Sean Escola, an American biotech founder and computational neuroscientist who recently moved to Berlin, I said he should come hang out for a few weeks here as he pondered his next venture in the “Neuro AI” space.

Sean joined us recently at the AI Campus to bounce ideas off of our Merantix Capital team and connect with the various parts of our Merantix Group — from AI solutions business Merantix Momentum to our Merantix Capital portfolio companies, like fellow biotech founder Greg Vladimer, as well as other residents here on campus.

This is a no-strings-attached offer we extend to seasoned serial entrepreneurs looking to build an AI venture in their domain. We are happy to let you tap into our ecosystem to help you on your way. You never know, maybe we’ll end up working on something new together via our venture studio at Merantix Capital, where we build and invest in category-defining application layer AI companies.

I recently sat down with Sean about his background and to ask him about his time working from the AI Campus. Our edited conversation is below.

Can you tell me a little bit about your professional journey?

I have somewhat of an unconventional background. I’m a clinical psychiatrist, and I led a computational neuroscience lab as an assistant professor at Columbia. I founded a biotech-meets-machine-learning company in San Francisco in 2017 called Herophilus, and we ultimately hired an external management team, although I stayed involved and was on the board. We had a business relationship with a company called Genentech for co-development for some drugs, but they actually got excited about our technology platform and decided in 2023 that they would acquire it.

The sale of the company was a good transition moment for my family. My wife is German and we wanted to be back in Germany. I moved to Berlin and tapped into my investor network, and I heard really good things about Merantix and the AI Campus. Then you and I had lunch and you suggested I come hang out on campus.

What are you looking to build next?

My last company was squarely in biotech, using a lot of machine learning applications to try to do data analysis of these large scale data sets that we were collecting for the discovery of new drugs. It’s modern techniques, but still a standard playbook.

A direction I’m now exploring, although it has an overlapping set of skills, is how we can use ideas about how the brain works as well as datasets that are collected from human or animal brains to train AI models in order to endow them with more human-like cognitive capabilities.

In other words, rather than using AI for neuroscience, it’s using neuroscience to build better AI.

What did interacting with the Merantix Group do for you?

I got a sense that you and Rasmus are really excited and helpful to founders who are familiar with one system and now thinking about doing something in a different system. For me, I understand the startup scene in the US but needed to understand more about what company formation looks like here.

There’s a utility to partnering with a team like Merantix who are very intimately familiar with the intricacies of building in Europe, and also at the same time recognize how to maintain efficiencies from the American system.

A major thing on my mind as a potential founder is access to good talent, and that’s really something that Merantix offers. You can really see it on campus, whether that’s through events or for instance the clear technical engineering competence at Momentum — and how they work with your portfolio companies.

You recently had a second child (congratulations!), but where are you at currently when it comes to building your next company?

We’ll have to see what happens in the next couple of months once I’m able to put more effort into trying to build something, but I would be very excited about potential partnering with the Merantix Capital team. From our time together, I could see firsthand how your relationship with your portfolio founders is supportive and quite collaborative, when it can sometimes be somewhat adversarial with investors, because your goals are often different.

Thanks for the chat Sean! You are certainly welcome on the AI Campus anytime, and we’re excited to see — and maybe be involved in — whatever you build next!

Are you a serial founder looking to build a category-defining venture on the application layer of AI? Get in touch, and come immerse yourself on our AI Campus like Sean did.

If you're an ambitious AI founder in 2025, the options have never been more varied or more confusing.

As someone who has been building and investing in AI since 2016 (and an entrepreneur myself), I spend a lot of time speaking with founders who are looking for their next move. They’re typically repeat founders with industry domain expertise interested in founding an AI-native venture and looking for the right partner to help get them there.

Sometimes we discuss Merantix Capital’s “founder-in-residence” program, where a founder explores multiple AI business cases over four to eight weeks with us, spends the next 2-4 months on validation ( building commercial traction), and if we both build conviction, we invest €1M for 15% preferred shares and 10% common shares (institutional co-founder equity).

To be quite frank, with multi-stage VCs pouring capital into pre-seed stage and “pre-idea” companies, there is mounting pressure on the studio model that makes it seem outdated. The reality is more nuanced. For the right type of founder, the venture studio model is not only viable, but the smartest path.

Caveat: I am biased. At Merantix Capital, our venture studio is core to our identity, through which we’ve helped create more than 10 amazing companies like Cambrium, Deltia, Ovom Care, and Libra. We also make investments in companies that were founded outside of our studio approach (about 50% of our latest fund).

Take it from me, over 20 years as a founder and eight years as an investor, I’ve seen trends in the early-stage ecosystem come and go, but the really successful founders have always recognized that how you start matters just as much as what you start.

Current market dynamics are dizzying

The early-stage AI market today is flooded with capital. Major VCs (especially those with large growth or crossover funds) have aggressively moved downstream, writing large checks at the earliest stages of company formation. In many cases, they're backing operators or repeat founders without a concrete idea, betting that these individuals will figure it out.

Why? In part, it’s a hedge. Late-stage investing has become more challenging, but the excitement around AI has created a sense of urgency to secure early access to promising talent. The big funds can easily afford to write off early bets as the cost of staying in the game.

This capital abundance brings its own issues. With more money chasing a relatively static pool of talent, founders are under pressure to scale fast and meet unrealistic expectations, often before they’ve validated a problem, let alone a tangible product or GTM.

What do venture studios do differently?

This is where the venture studio model (and I would argue our venture studio!) stands apart. Instead of betting on potential, we are a partner in the core foundation of a successful venture: early validation, commercial traction, capital efficiency, and team quality.

It’s true, we write smaller checks than a mega-fund, and our founders may take slightly more dilution upfront. We won’t argue otherwise. But we offer something large funds can’t: deep support at the earliest stages from an AI ecosystem that includes corporate design partners, domain experts, and an in-house talent pipeline centered at our AI Campus in Berlin and the AI Hub in London.

Many of our portfolio companies have paying clients at or soon after incorporation—often secured through our ecosystem. Four of our last five studio startups, for example, reached €500K in revenue within the first six months. Three also went from pre-seed to seed in under 10 months.

That level of early traction simply isn’t possible without trust and access, two things the Merantix platform offers in spades.

Efficiency is the edge

Capital efficiency is more important than ever these days. Startups without early validation often overbuild. That means bigger teams and more burn just to get in front of customers. Our approach flips that script. With early access to design partners and strategic customers, our startups can iterate faster, avoid premature scaling, and stretch €1M in ways others need €3-5M.

Other investors might want a fully-formed team, but we’re comfortable starting with a strong founder and co-building the right team from there (including co-founder matching through our network). That flexibility means we can bring in the right people at the right time without overstaffing.

When we put out a call for specific profiles on these teams, we receive hundreds of applications within days. That talent density is a force multiplier, and one of the key reasons our companies punch above their weight.

Who is the venture studio model right for?

The plain truth is that the studio model isn’t for everyone. It’s best suited to founders who:

• Are repeat entrepreneurs and understand the complexity of building from scratch

• Have deep operational experience and want to lead but not go it alone

• Bring technical expertise but need commercial or regulatory co-founders in complex sectors

The venture studio model is, in short, best for founders who are long-term thinkers. Those are usually the best founders! They understand that speed, talent, and traction matter more than retaining an extra few points of equity early on. Not to mention that higher traction in the beginning typically means less dilution in later rounds for founders.

Moving beyond the buzz

Why did I feel the need to write this? I don’t want to sound defensive or over-advertise our venture model. It isn’t for everyone and, like I said, half of our fund is geared towards “non-incubations” too.

But it also has bugged me that “venture studio” has become a dirty word in VC, because I think it’s a great model that produces huge results.

There’s a lot of noise in the AI world right now. Every startup is "agentic” and every founder is a visionary. But building enduring companies still requires the hard stuff: talking to users, recruiting great people, and making products people actually want.

That’s what we help founders do. Not through hand-holding, but through actual partnerships and a trusted brand. Our value is acceleration through talent and traction.

It’s no secret that “Agentic” has been the AI buzzword of 2025. The problem is, most of the conversation still feels abstract. It's hard to know what’s real, what’s working, and what’s just noise.

In this piece, I want to cut through that. I’ll walk through examples of agentic AI that we find most interesting at Merantix Capital — from financial services to science to healthcare — and share a few thoughts on what makes the best systems stand out.

What is Agentic AI

For the uninitiated, let's briefly define Agentic AI: it’s essentially an AI system that can independently assess a problem, plan a series of actions, execute that plan across tools, and learn from the outcome.

But not everything branded as "agentic AI" today actually earns the title.

Cutting Through the Noise

The term “agentic” has been stretched so far it covers everything from basic automation tools to genuinely autonomous systems. Buyers that think they are getting an AI agent can sometimes end up with a narrow tool with no feedback loop, no memory, and no ability to improve. (Microsoft Copilot, for instance, is a useful service but not yet capable of managing complex workflows without human oversight).

At Merantix Capital, we look for agentic systems that are solving painful, high-friction problems inside the industries that run the world. The teams that matter combine technical firepower with deep domain expertise. They build agents that are not just clever, but dependable, battle-tested and capable of delivering compounding results.

Payments infrastructure, scientific research, healthcare delivery, cybersecurity — these are just a few of the frontiers where agentic AI will prove itself. I want to provide a few examples of innovative solutions in these spaces.

Payments

Payments is one of the first sectors where agentic AI is stepping beyond mere recommendation into real-world action. Traditionally, even the most capable AI systems were bottlenecked: they could suggest transactions but not actually execute them. The problem was not intelligence, it was infrastructure. Existing payment systems were built for human users, not autonomous agents.

Skyfire, Nevermined, and Payman are three startups that represent different but complementary solutions to this bottleneck.

Skyfire built a payment network designed specifically for AI agents to spend money safely. Rather than tying agents directly to a user's bank account, Skyfire issues each agent a wallet ID funded via traditional methods like ACH or wire, but all transactions conducted by agents happen in digital assets like USDC. This architecture allows businesses to tightly control spending through real-time dashboards while enabling machine-speed execution. In beta deployments, enterprises like car parts firm Denso used Skyfire to automate procurement payments, allowing agents to complete supply chain transactions without human intervention.

Nevermined, meanwhile, is building the "PayPal for AI agents," focusing on AI-to-AI commerce. It allows agents to pay one another for services autonomously, using crypto rails optimized for dynamic pricing, metered usage, and trustless transactions. Their approach emphasizes decentralized financial infrastructure, helping agents bypass the rigid, account-based systems that slow down human-driven commerce.

Payman tackles the other side of the problem: letting AI agents pay humans directly. Payman’s platform links agents to traditional banking systems and supports payments not only in USDC but also through fiat rails like ACH and wires. This makes it one of the first to bridge agents into the human economy, enabling scenarios where an AI agent could automatically hire a freelancer, monitor task completion, and release payment without manual coordination. Unlike Skyfire and Nevermined, which still transact primarily in digital assets, Payman enables actual bank transfers alongside crypto.

Most early agentic payment systems rely on digital assets like USDC or stablecoins because traditional banking infrastructure is not yet designed for autonomous, machine-initiated transactions. While Payman has made inroads into traditional rails, the broader financial ecosystem is still catching up.

This emerging layer of infrastructure, including wallets, safeguards and programmable transaction limits hints at what broader autonomous commerce could look like. As the rails mature, there is enormous potential for founders to build orchestration layers, audit and compliance tools, fraud prevention systems, and seamless agent-to-agent marketplaces on top.

Science and Healthcare

Scientific research and healthcare both suffer from massive knowledge bottlenecks: too many papers to read, too many variables to test, too many patients to monitor. Traditional workflows are slow, heavily manual, and ripe for transformation by agentic systems.

AI Scientist-v2 is an agentic system that performs the entire scientific method autonomously. It generates hypotheses, designs and simulates experiments, evaluates results, and iterates toward better answers. Then, it writes up those findings as full academic papers.

Introduced in ‘The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search’, by Yutaro Yamada et al in April 2025, this model contains a remarkable leap forward both AI and science: a fully AI-generated paper with no human ghostwriting that passed peer review. That’s a first in scientific publishing.

Here’s how it works:

1. The system explores multiple experimental pathways in parallel, simulating experiments based on probabilistic models and refining hypotheses over time.

2. Each hypothesis is scored for novelty and plausibility.

3. Promising results are bundled into a structured research paper using a modular writing agent trained on scientific publication norms.

4. Then, the paper is submitted for blind peer review. This time it got in.

This is autonomous research. AI that contributes new knowledge, not just insights. And it raises critical questions not just about what AI can do, but what roles remain essential for humans in the scientific process.

Another important example is Causaly, a biomedical research company that developed an agentic AI platform autonomously planning and executing multi-step research workflows by drawing from a 500 million–fact biomedical knowledge graph and external scientific literature.

Causaly’s agentic system reportedly prioritizes relevant evidence, updates findings dynamically as new data emerges and traces every insight back to its source to maintain scientific transparency.

In healthcare, Hippocratic AI, co-founded by Munjal Shah alongside experts from Johns Hopkins, Stanford, and Google, is tackling the shortage of frontline medical staff.

Hippocratic AI’s platform trains agentic healthcare assistants to handle non-diagnostic tasks like post-discharge follow-ups, chronic disease coaching, and medication reminders, while maintaining a rigorous safety-first approach validated through clinical trials involving over 5,000 nurses and 500 physicians. By late 2024, Hippocratic AI reported that its agents had achieved safety parity with trained nurses and pharmacists, formed partnerships with major health systems like Memorial Hermann and Universal Health Services.

These examples show that agentic systems aren’t just accelerating research or augmenting healthcare, they are starting to expand the boundaries of what is possible by performing tasks that were once considered uniquely human.

That said, there are still major opportunities ahead. Today, most scientific and healthcare agentic systems operate mainly in digital environments, working with existing data rather than interacting directly with the physical world.

This is partly because it’s easier to coordinate digital tools than to manage real-world variables like lab equipment or patient care. However, there is a huge need for experts to build agentic systems that can also automate wet lab experiments, manage clinical trials, and coordinate ongoing patient care — areas where sensitivity to real-world context, safety, and adaptability are essential.

Cybersecurity

Cybersecurity is a race where traditional defenses are falling behind. New vulnerabilities emerge daily, software evolves constantly, and attackers adapt faster than human teams can respond. Quarterly penetration tests, static firewalls, and reactive patching are no longer enough. The next wave of defense needs systems that think, move, and adapt on their own.

Terra Security deploys fleets of autonomous agents that continuously probe enterprise environments like web apps, APIs, and cloud assets without waiting for human instructions. Each agent specializes in a class of exploits such as SQL injection, authentication bypass, or privilege escalation, and plans new attacks dynamically as it learns about the target’s structure.

This shifts cybersecurity from point-in-time testing to continuous, self-directed red teaming. Instead of waiting months for a vulnerability report, companies receive real-time findings prioritized by exploitability and business risk.

Terra uses a hybrid model where human analysts verify AI-generated findings to ensure accuracy and relevance. This combination of machine scale with human oversight addresses one of the biggest weaknesses in traditional penetration testing: inconsistent accuracy and lack of operational context.

Despite these advances, cybersecurity remains wide open for agentic innovation. Most systems today focus on offense. Defensive agentic systems, capable of predicting threats, patching vulnerabilities, and neutralizing attacks without human help, are still rare. There is enormous opportunity for security-native founders to build the next generation of autonomous, self-healing defenses. (In our own portfolio, check out Company Shield).

The best is yet to come

The best agentic projects rise above the noise because they deliver genuine autonomy where others still depend on human intervention. The breakthroughs in payments, research, healthcare, and cybersecurity show that when agentic systems are built for real-world complexity, they do not just streamline workflows, they open up entirely new ways of operating (check out our recent Substack with LogicStarAI’s founder for another example in the engineering space).

Founders who combine technical strength with deep industry knowledge, and who focus relentlessly on solving high-friction problems, will define the next era. The future belongs to those who can turn agentic AI from a buzzword into real-world performance.

As always - if you are ideating or building in this space, we should talk!

For an in-depth pulse check on the Agentic AI hype, be sure to join our AI Breakfast on this topic, being held this coming Thursday, June 12th on the Merantix AI Campus: https://lu.ma/7oazc49z

Berlin / Hamburg – Today, Jungheinrich and Merantix announced a pioneering strategic partnership. The aim of the cooperation is to promote Artificial Intelligence (AI) based applications in the industry.

As part of the partnership, Jungheinrich is investing five million euros in Merantix Capital's AI fund. The fund invests in AI-native start-ups building impactful, scalable solutions that address meaningful challenges in the industrial environment, as well as in health care, finance, manufacturing, and other verticals.

Dr Lars Brzoska, CEO of Jungheinrich AG, says: “Jungheinrich has extensive know-how in the field of material handling and engineering. With Merantix, we are gaining a strong partner with in-depth expertise and strong competence in building AI-based business models. The partnership will enable us to create valuable synergies and work together to strengthen AI applications in the industrial environment.”

The strategic partnership will focus on AI investment opportunities in Europe, particularly in key areas such as robotics, sustainability, enterprise AI, energy, and logistics. Possible concrete use cases for the collaboration include solutions for analysing used lithium-ion batteries or for warehouse simulation with the aim of optimising material flows.

“AI-native business models are changing the industry. However, their success depends largely on their successful integration into existing structures,” says Dr Rasmus Rothe, co-founder and General Partner of Merantix Capital. “At Merantix, we know the requirements for scaling AI companies. As an industry partner, Jungheinrich brings the necessary experience and openness for innovation.”

The partnership with Merantix complements Jungheinrich's existing activities in the area of corporate venture capital. It follows the investment in the High-Tech Gründerfonds (HTGF) and is part of Jungheinrich’s new strategic direction of systematically developing further business models outside its core business via the recently founded subsidiary Uplift Ventures. Additional fund investments are planned in the future to give Jungheinrich broader access to the start-up ecosystem and further expand the international reach of its venturing activities.

The partnership also represents Merantix Capital’s deep ties with key industry partners in its effort to build a bridge between corporations and the AI startup ecosystem, helping to drive the kind of necessary AI adoption in Europe that can only happen through close collaboration and knowledge sharing.

In addition to the investment in the Merantix Capital fund, the partnership includes cooperation with other units of the Merantix Group, including AI solutions provider Merantix Momentum, as well as exchange with the Merantix community at the AI Campus in Berlin, Europe's largest AI centre with around 80 companies and over 250 events annually.

If you have any queries, please contact:

Dr Benedikt Nufer, Spokesman, Jungheinrich AG

Tel.: +49 40 6948 3489

Mobile: +49 151 2779 1245

benedikt.nufer@jungheinrich.de

Steven Perlberg, Head of Communications, Merantix Capital

steven@merantix.com

About Jungheinrich:

As one of the world’s leading providers of material handling solutions, Jungheinrich has been advancing the development of innovative and sustainable products and solutions for material flows for more than 70 years. As a pioneer in the sector, the family-owned listed business is committed to creating the warehouse of the future. In the 2024 financial year, Jungheinrich and its workforce of around 21,000 employees generated revenue of €5.4 billion. The global network comprises 12 production plants and service and sales companies in 42 countries. The share is listed on the MDAX.

About Merantix:

Merantix is Europe’s leading AI group, focusing on investing, building, and growing AI. The Merantix group includes Merantix Capital, an early-stage venture capital firm working with founders to bring AI into real-world applications in sectors ranging from healthcare to manufacturing to life sciences and enterprise. Founded in 2016, Merantix’s other initiatives together make up the largest AI community in Europe, including AI solutions business Merantix Momentum and the Merantix AI Campus in Berlin Mitte, a coworking and events platform with more than 80 companies and 200+ annual events.

Last month, Merantix Capital had the pleasure of hosting a great AI pitch competition at the annual START Summit in St. Gallen, Switzerland.

We heard from six great early-stage AI ventures, but one really stood out to our panel of judges: LogicStar AI, which is building autonomous AI agents that solve one of the biggest bottlenecks in software development (the manual and slow task of maintaining and debugging production code). The company recently announced its €2.9 million in pre-seed funding round, led by Northzone.

I had the opportunity to sit down with LogicStar AI CEO and co-founder Boris Paskalev after his successful pitch to talk about how he’s creating “Silicon developers” and the future of agentic AI. A seasoned founder, Paskalev’s last venture DeepCode was acquired by Snyk.

Adrian: Congrats on winning our pitch competition. Agentic AI is obviously a hot area right now. Tell us a bit about your core vision and how you define an agent.

Boris: Our internal definition of an agent is an entity, ultimately a “Silicon developer,” that fully autonomously produces tasks and calls large language models. We distinguish between agents and tools. Agents can call other agents or tools and work together to produce what we want. Ultimately, it’s about freeing engineers from the laborious tasks of maintaining code—investigating, reproducing, and fixing issues with high-fidelity, production-ready code.

So it’s about allowing engineers to focus on creativity?

Exactly. A lot of agentic AI tools talk about freeing up engineers for more meaningful work, and I agree. It’s already quite autonomous. The question is: what percentage of the backlog can be fixed autonomously?

That kind of automation frees up a significant number of engineers in larger organizations to focus on building new things. As our “junior Silicon developer” gains experience, it’ll take on more things like fixing documentation, refactoring code, and validating human work. But the biggest value today is automating a part of the backlog without engineers lifting a finger and doing it much faster.

How much faster?

From what we’ve heard consistently, it takes engineers between one hour and a day to fix a bug, on average maybe four hours. We can fix it in seconds. That’s the real value. It’s about building self-healing applications and fixing issues before customers even notice.

Do you see this ever replacing developers?

No, we’re not claiming we’ll replace developers. Complicated stuff will still need a human. But our goal is to take away a large portion of the work they don’t want to do. One of our key differentiators is that we’re not just building a co-development tool.

Today, most tools that are called “agents” expect constant human input. You prompt, you review, you discard or tweak 80–90% of the output. That’s just not scalable. The return on investment isn’t there. You spend more time reviewing than just fixing it yourself.

Are there particular industries where LogicStar AI works best?

Honestly, it’s across the board. We’ve seen it work from big banks to small startups, even open-source libraries. The main limitation right now is language. We only support Python. JavaScript and TypeScript are next.

The real challenge with adding a new language isn’t technical. It’s building the benchmarks to test the agents properly and ensure they improve over time. So when talking to folks on our waitlist or potential design partners, the main question is: “Do you have a Python app?

You’ve founded before. What's your approach to company building, based on your past experience?

Talk to customers as early as possible. Right now, we’re onboarding design partners and building out the initial integration. We want them to shape the product. The focus is building a strong product, early go-to-market and visuals that show the value clearly before we scale up sales and support teams.

What’s the benefit of building in Europe?

Honestly, it’s the deep tech community. ETH Zurich and INSAIT have incredible researchers. My co-founders are top experts in program analysis and LLMs. We’re not lacking in talent at all.

We’re also building out frontend teams in Bulgaria, where we have strong ties and can recruit great engineers. But customer support and sales will likely be U.S.-based, since that’s where a lot of our early adopters are.

What’s your LLM strategy?

We’re model agnostic. We use Claude, GPT-4 Mini, Gemini 2.5. Whichever works best for a given task. Different agents perform better with different models.

We’ve already got 12 agents working across different steps, and each uses a different model depending on what’s needed. That flexibility is key. As models improve, our product improves. We also care a lot about cost-efficiency. We want to build a sustainable business, not one that loses money hoping to scale later.

How did winning the pitch competition at START help you?

It gave us visibility. But really, it reinforced our mission: to build the first silicon maintenance engineer out there that frees up engineers.

Thanks very much, Boris! Good luck to you and your team, and congratulations again!

Today we have some great news from Libra, our stellar legal AI company which has just announced its latest update: Libra v2.

I’ve been super impressed with the Libra team under the leadership of founder Viktor von Essen, a trained lawyer who joined us at Merantix Capital as a founder-in-residence in 2023.

Libra’s mission is straightforward but challenging: becoming the go-to AI service for lawyers when it comes performing the more menial and time-consuming tasks of legal work, from research to contract review.

With Libra v2, the company has taken a big step forward. I recently sat down with Henrik Volkmann, Libra’s Director of Growth who joined the company in March, to hear about v2 and Libra’s roadmap for the future.

Adrian: Can you tell me a little about Libra v2? What’s the vision behind this update and how does it reflect what you’re hearing from the lawyers on the platform?

Henrik: Our goal is to redefine efficiency and precision in legal work. We've achieved this by integrating advanced AI, a sleek user interface, and practical tools, including our new Libra AI, which offers direct interaction and intelligent, agent-based reasoning tailored to the legal field. We're also rolling out a suite of new features and significantly improving the overall user experience. This update is directly driven by feedback from lawyers, ensuring that Libra v2 supports every aspect of their workflow, from initial research to final contract review. Ultimately, we want legal professionals to focus on what they do best: providing exceptional legal counsel

.Tell me about Noxtua, a specialized LLM for legal. What makes it uniquely suited for legal work compared to general-purpose models?

Noxtua is a specialized LLM for legal tasks, developed by our close partner, Xayn. It's fine-tuned using court decisions and legal documents like contracts, and it's updated monthly to incorporate new court decisions. A key advantage of Noxtua is its accuracy; it finds real court decisions and avoids hallucinating, providing links to the published decisions for each quote. Additionally, Noxtua excels at using legal language and writing legal documents in a manner preferred by lawyers.

What kind of use cases does Libra v2 excel at?
It's often tasks such as summarizing information or creating a first draft, but using our new Review feature for analyzing legal documents against internal standards and extracting relevant information from large documents is definitely a “killer feature.” This allows lawyers to quickly digest complex information and ensure compliance.

What do you think the future of AI looks like in the legal world, and how will Libra be a part of that?

We see AI becoming increasingly integrated into all facets of legal work, from initial research and case preparation to contract negotiation and even litigation support. Libra aims to be at the forefront of this transformation, continuously evolving to meet the changing needs of the legal profession. In the future, we will crack even longer and more complicated workflows and processes of legal work, automating and streamlining even the most intricate tasks. We envision a future where AI augments, rather than replaces, human expertise, allowing lawyers to focus on the strategic and nuanced aspects of their work that require human judgment and empathy.

What has surprised you about how Libra is being used so far by lawyers?

A surprising finding is that not only young lawyers but also older, more experienced lawyers are showing a great deal of interest and enthusiasm for Libra, recognizing its potential to streamline their workflows and enhance their practice.

Privacy concerns must be at the forefront of lawyers’ minds, particularly interacting with AI. How do you make sure that Libra is compliant and safe to use?

Regarding compliance and safety, Libra is compliant with GDPR and BRAO, and holds an ISO 27001:2022 certification. We meet the very strict compliance requirements that German lawyers have. To ensure the secure processing of sensitive client data within Libra, we have established comprehensive non-disclosure agreements with subprocessors such as Microsoft and Deutsche Telekom. German lawyers are required to have these agreements in place with us and our subprocessors. This allows us to compliantly process sensitive personal data.

With more than 3,000 lawyers already using Libra, where do you see the biggest opportunities for further growth? What’s the roadmap ahead?

The biggest opportunity for further growth is right in front of us. Most law firms are not yet using AI tools in their practice, and we often have the chance to be the first to demonstrate the possibilities. We're also engaging with many corporate legal departments. In the near future, we're focused on building more relevant features for legal professionals that will simplify their work and free up their time for actual legal tasks. We see significant potential in integrating more knowledge databases to enhance legal research, and also in further integrating Libra into the software lawyers already use, such as Microsoft Word and Outlook, as well as their document management systems.

Can you tell me a little about your background and your new role at Libra?

I am a fully qualified lawyer by training but have been a tech enthusiast since childhood, enjoying building computers and solving technical problems. At university, I founded a student-run non-profit to provide law students with experience in legal technology. Over the last years I gained experience from other startups and also worked in international law firms like Hogan Lovells and Noerr. At Libra, I am the Director of Growth and responsible for developing our customer base, which also involves a lot of work on the product side. I believe it's very important to have a legal background when selling to lawyers.

A bit of a self-serving question, but here goes: How does our Merantix ecosystem help you in your career as well as your work so far at Libra?

Being part of the Merantix ecosystem has been absolutely fantastic! The Merantix AI Campus is a vibrant hub, and the network it provides is truly incredible. I've had the chance to connect with so many brilliant colleagues from other AI startups, and the knowledge exchange has been phenomenal. We're not just talking about technical details either – we're diving into go-to-market strategies and sharing valuable insights that are directly impacting our growth at Libra. And let's not forget the little things that make a difference – I've even perfected my latte art skills using the awesome coffee machine! It's a supportive and inspiring environment that's definitely fueling our success.

Thanks very much for the chat and congrats on Libra v2!

In recent conversations with founders and investors alike, one theme keeps resurfacing: in today's AI landscape, switching costs are dangerously low. As someone who spends every day thinking about AI innovation and its long-term trajectories, I believe this is one of the most underappreciated risks (and biggest opportunities) in building enduring companies.

In venture investing, you're betting on companies over 5-10+ years. Traction today is meaningless if users can easily leave tomorrow. If a startup loses momentum because customers switch, regaining it is extremely difficult. Commoditization pressure is real, especially as many foundation models and AI tools increasingly resemble commodities (I highly recommend this article by my good friend Rob Toews.)

In this substack I wanted to outline Merantix Capital’s approach to dealing with switching costs, why they matter more than ever, how we believe startups can build defensibility, and how emerging technologies like continual learning might change the game.

In short, the goal of an AI founder should be to create switching costs so high that replacing their product feels like replacing a well-trained, high-performing employee -in other words, very difficult.

Why Switching Costs Are So Low in AI

• APIs and models are interchangeable. Companies can swap model providers almost overnight. In many cases, swapping from one LLM provider to another is as simple as updating an API endpoint. I see this in my own day to day: I switch tools as soon as a slightly better, cheaper, or faster option appears.

• The speed of innovation is relentless. New models and techniques are appearing every few months. Even companies with strong technical teams can be outpaced if they rely purely on product speed without building underlying defensibility.

• B2B switching happens faster than consumer switching. Business users are rational, technical, and highly cost-driven. Unlike consumer behavior, which is often sticky due to habits, brand perception, or inertia, business decisions are made on ROI, performance, and cost…

• Commoditization is accelerating. Models today largely deliver similar outputs, and continual learning is not yet widely adopted. This makes it very easy for users to move from one provider to another without any real switching pain.

• Low integration depth. Many AI applications layer thinly on top of models without deep integrations into broader systems (e.g., ERP, EHR), making it easier for users to switch when something marginally better arrives.

How to build and evaluate high switching costs

To build truly durable companies in AI, founders must think deliberately about switching costs from day one.

We find it useful to think of switching as compounding layers.

The deeper you move a customer down these layer, the harder it becomes for users to switch away:

Each layer deeper on this chart compounds the stickiness of your product.

The Emerging Frontier: Memory, Continual Learning, and Durable Moats

Today’s LLMs are powerful but fundamentally static: they learn during training and then stop adapting. True human-level AI requires models that can learn continuously throughout their deployment lifecycle.

Several workaround techniques exist today:

• Fine-tuning: Helps update models post-deployment, but is batch-based, expensive, and limited.

• Retrieval-augmented generation (RAG): External memory approaches are useful but face scaling and latency challenges.

• In-context learning: Allows models to "simulate" adaptation during inference but without true weight updates.

However, none of these approaches solve the deeper problem of building durable, evolving AI models.

Continual learning, where models update their own weights on the fly without retraining from scratch, represents a profound shift. In his article, Toews argues that continual learning will, “unleash AI’s full potential to power hyperpersonalized and hypersticky AI products.”

In other words, imagine an AI legal assistant that becomes better at drafting documents in your firm's style over time. Or a recruitment AI that learns your exact criteria for screening candidates without explicit reprogramming. Models would develop deep, personalized knowledge that is hard to replicate elsewhere, creating switching costs akin to replacing a well-trained, high-performing employee.

Companies like Writer and Sakana AI are already experimenting with architectures that enable this self-evolving behavior. These breakthroughs hint at the future direction of defensibility in AI products.

How Founders Should Think About Switching Costs from Day One

Switching costs aren’t a "Series B problem." They must be part of the early design and go-to-market strategy:

• Move fast but design for depth: Speed to market is crucial, but not at the expense of defensibility. Focus on integrations, memory, and workflow entrenchment.

• Be intentional about data: Find ways to own, structure, or personalize user data.

• Audit your defensibility constantly: Revisit the checklist regularly as the product evolves.

• Think ahead to continual learning: Even if not feasible today, position your architecture for future personalization and adaptation.

Conclusion

In AI, defensibility is no longer just about building technically superior products. It’s about durability: how resilient your customer base is in the face of rapid technological change.

That's why I prioritize investments in startups that integrate deeply into critical workflows, prioritize user trust, create strong switching costs, and build toward the future of dynamically learning AI systems.

If you're building a company that understands this challenge and is actively working on durable moats in AI, I'd love to hear from you.

Good health is a privilege, not a guarantee.

That’s why healthcare is unique as an industry that we will all interact with throughout our lives, whether we like it or not.

Beyond the sisyphean task of day-to-day patient care, healthcare is a creaking sector of ancient IT systems, poor tech integration, mountains of siloed data and manual processes crying out for automation.

It’s also one of the most exciting for innovative AI applications, as technology is finally maturing to a point where it can make a meaningful impact for patients, hospitals and beyond.

In this post, I want to highlight some use cases I think are overdue for their time in the AI sunshine. Thanks also to my colleague Bertram Weiss, VP Health at Merantix Momentum, for providing his input.

Hospital Logistics and Bed Management

Inefficiencies in bed allocation, medication supply chains, and staff scheduling all lead to wasted resources and delayed patient care.

Bed availability is one of many operational pain points for hospitals. But where administrators see a chaotic mess of complex data and poorly integrated systems, AI can streamline and provide actionable insights, as a 2019 Harvard Business Review found.

• A John Hopkins Hospital implemented an AI-powered control center that improved bed assignment efficiency by 30%, reduced the need to keep surgery patients in recovery rooms longer than necessary by 80%, and cut the wait time for beds for incoming emergency room patients by 20%.

AI is also being trialled in transforming predictive inventory control, particularly for medications and critical medical supplies. AI-powered time series forecasting models, including ARIMA (AutoRegressive Integrated Moving Average) and LSTMs (Long Short-Term Memory networks), anticipate fluctuations in medication demand, helping hospitals prevent stockouts and avoid costly overstocking.

A study published in Acta Logistica evaluated the performance of ARIMA and LSTM models in forecasting the demand for five commonly supplied medicines in Rwanda's public health system. The LSTM model demonstrated superior accuracy, achieving a 95.2% performance rate in demand prediction, indicating its potential to significantly reduce stockouts and overstocking.

If you’ve ever tried to get a doctor to give you a firm answer on whether they can attend a future event, you will know how astoundingly chaotic hospital staffing methods are. Days off have to be booked months in advance and understaffing means those precious days off for overworked medical staff can simply disappear into thin air.

Fortunately, AI-driven reinforcement learning algorithms are improving hospital staffing models, dynamically adjusting workforce deployment based on expected patient inflows, peak hours, and emergency events. Bioengineering Basel’s 2024 report demonstrated that AI applications could predict patient admissions and discharges, allowing for dynamic staffing adjustments. This proactive approach helps in maintaining optimal nurse-to-patient ratios, thereby improving patient care and reducing staff burnout. The report also suggested that this data could help improve patient experience by giving accurate information about appointment times, wait periods and treatment schedules.

AI-driven scheduling is also cutting costs and improving efficiency in the operating room, according to a 2024 report from the Fort Lauderdale Eye Institute.

• A machine learning algorithm designed to optimize operating room booking times led to a 21% reduction in nursing overtime, resulting in significant cost savings.

• Over a three-year period, the reduction in overtime hours translated to $469,000 in savings, highlighting AI’s potential to alleviate staff burnout and improve operational efficiency.

These improvements aren’t just reducing costs—they are fundamentally reshaping hospital management by ensuring that every resource, from beds to staff to equipment, is utilized at peak efficiency, directly improving patient outcomes and hospital profitability.

AI-Powered Medical Scribes

Doctors spend a third of their working hours filling in paperwork, a Hett Insights report found. It doesn’t stop when they clock off either, with some spending nearly 5 hours of their personal time per week wrestling with documentation. This is a staggeringly poor use of time that contributes towards burnout and something that can easily be taken over by AI scribes.

AI scribes leverage automatic speech recognition (ASR), natural language processing (NLP), and large language models (LLMs) to extract key information from conversations in real time. AI-generated notes follow standardized formats such as SOAP (Subjective, Objective, Assessment, Plan) and BIRP (Behavior, Intervention, Response, Plan), ensuring consistency and accuracy, according to an NEJM Catalyst study.

A pilot study on AI scribes in dermatology found that they significantly reduced documentation time, enabling doctors to focus more on patient interactions instead of administrative tasks. The American Medical Association (AMA) reported that AI scribes could save doctors one hour on average, drastically improving efficiency and reducing burnout among healthcare professionals.

AI scribes aren’t just a helpful notetaker for doctors making their rounds, they’re also a valuable source of health record and insurance data that can be integrated into future multimodal AI.

Multimodal AI will allow for data from text, such as doctor’s notes and EHRs, to be examined along with images and time series data like X-rays, MRIs and monitoring data from wearables and hospital equipment. AI will assist doctors by quickly synthesizing complex patient histories and identifying correlations that would take a human much longer to process.

Scribes have huge potential to cut back on paperwork and increase the time doctors share with their patients. Any opportunity to automate bureaucracy faced by clinical staff should be taken up by innovators with the dual goal of helping to increase patient rate and generate a valuable data pool from the associated claims data that can help with personalized patient interventions, disease forecasting and working towards a 360 degree multimodal approach to patient care.

AI for Epidemic Detection and Pandemic Preparedness

The global spread of COVID-19 laid bare the grim reality of pandemic preparedness. Thankfully, the worst of the coronavirus is in the rearview mirror, but experts warn that we are far from prepared for the next global outbreak.

The bad news is that we don’t know when the next one will strike. The good news is that AI might be our best chance to minimise the impact when the time comes.

A February 2025 study from the University of Oxford highlights the transformative role of AI in predicting, detecting, and containing pandemics before they escalate. The study found that AI-powered epidemiological modeling, genomic tracking, and real-time surveillance are essential tools in strengthening global pandemic preparedness:

• Improving current models of disease spread, aiming to make modelling more robust, accurate and realistic.

• Pinpointing areas of high-transmission potential, helping ensure limited healthcare resources can be allocated in the most efficient possible way.

• Improving genetic data in disease surveillance, ultimately accelerating vaccine development and the identification of new variants. Real-time surveillance using genomics such as PCR means wastewater can now be monitored as one of the earliest sources for detecting pathogens, something that was successfully implemented during the pandemic.

• Helping determine the properties of new pathogens, predict their traits and identify whether cross species jumps are likely.

• Predicting which new variants of already-circulating pathogens – such as SARS-CoV-2 and influenza viruses – might arise, and which treatments and vaccines are best in reducing their impact.

• Possible AI-aided integration of population-level data with data from individual-level sources – including wearable technologies such as heart rate and step counts – to better detect and monitor outbreaks.

• AI can create a new interface between the highly technical science and healthcare professionals with limited training, improving capacity in settings that need these tools the most. This can bridge between staff that went into the sector for the ‘human interaction’ and help speed up the relatively slow adoption of technology in healthcare.

AI is also proving essential in vaccine development, where it accelerates genomic analysis and drug repurposing, cutting traditional vaccine research timelines by up to 50%. Pharmaceutical firms such as Moderna and Pfizer are actively integrating AI into their R&D processes, allowing for faster identification of high-risk mutations and more precise vaccine targeting.

Epidemiologists are quick to remind us that the next pandemic won’t politely wait around the corner until we’re ready: it’s only a matter of time before the next infectious disease takes hold. The time for AI adoption must be between pandemics, not during one.

In my book, that’s a pretty urgent call to action. For ambitious builders and healthcare experts looking to create tools that will help manage the next pandemic, yesterday was the best time to start, today is the next best opportunity.

AI-Powered Wearables and Health Monitoring

AI-enabled wearable devices are revolutionizing real-time health monitoring, providing continuous, non-invasive insights into patient well-being. These wearables offer early detection of health conditions, personalized interventions, and enhanced patient care. Recent research highlights their growing role in cardiovascular health, human activity recognition, and intensive care monitoring.

A 2023 study showed that wearable ECG devices demonstrate high efficiency in processing single-lead ECGs. The research shows that the devices can deliver near real-time cardiac assessments in 33 to 35.7 seconds, ensuring faster diagnosis and early intervention. These ECGs improve arrhythmia detection and reduce reliance on manual interpretation, helping healthcare professionals manage cardiac patients more effectively.

A 2025 study published in npj Digital Medicine found that a deep learning model trained on a dataset of over 150,000 individuals found associations with 1,600 diseases, outperforming traditional ECGs:

“When compared to standard ECG intervals, the autoencoder-based latent space models reveal a substantially greater number of associations, suggesting that the rich representations afforded by the architectural complexity of the deep learning model provide more information about disease status than routinely ascertained ECG intervals.”

The paper also notes that consumer-based ECG devices could enable large-scale, remote disease detection.

This powerful predictive is backed up by another study from the American Heart Foundation which reported that devices in the ‘Multisensor Monitoring in Congestive Heart Failure (MUSIC) study’ developed and validated a personalized algorithm to predict acute decompensation in 543 patients with heart failure. The wearable device predicted impending hospitalization with 76% to 88% sensitivity at a median of 6.5 days before hospital admission. That’s a huge window for preventative action.

These wearables are also a sign of another trend: building systems outside of traditional healthcare infrastructure.

Traditional infrastructure, such as hospitals and surgeries, often use outdated information systems such as EHR/EMR and in some cases wouldn’t be able to integrate live data from wearable devices. That’s why some innovators are trying to reinvent healthcare end-to-end: to allow the AI to work to its fullest potential and rethink the experience entirely. It’s also another reason why it is easier to set up this independent infrastructure in prevention.

Just look at Spotify co-founder Daniel Ek’s new AI-powered full-body scan and preventative healthcare platform, Neko Health. The firm’s clinics use AI and non-invasive sensor technology to generate millions of data points across cardiovascular, metabolic, and dermatological health indicators. Patients receive comprehensive AI-assisted insights into potential early-stage health risks such as cancer and cardiovascular disease.

They say it’s the next step towards preventative health, although I think there are some fairly large structural and incentive challenges in the health industry before doctors start prescribing jogging and earning from healthy patients rather than sick ones. That said, a pilot test is underway in Kinzig Tal doing just that.

Either way, wearables have enormous potential to improve the doctor’s ability to diagnose and treat patient’s illnesses, in some cases allowing for preventative action. On top of this, new AI tools could connect siloed databases and insurance code to improve tracking in wearables and immediately send alarming data to doctors. We will also see new kinds of sensors that are able to monitor different variables. This unprecedented availability of patient data should be mined to bring real-time nowcasting into healthcare and make trend-spotting, diagnosis and treatment as dynamic as they can be.

An industry more than ripe for AI

Healthcare isn’t just another industry—it’s a system we all rely on that remains crippled by inefficiencies and missed opportunities.

AI presents a rare opportunity to rebuild critical processes from the ground up, not just optimize broken ones. For founders and innovators, this isn’t about incremental improvements; it’s about reshaping how care is delivered, predicted, and personalized around the world.

The first stages of preventative health are finally within reach, with AI wearables catching cardiac events before they happen and predictive models identifying disease risks early. Hospitals are no longer bound by guesswork, as machine learning streamlines logistics, staffing, and resource allocation. Meanwhile, AI-driven automation is freeing doctors from paperwork, returning them to patient care where they belong.

The challenge now is execution. Legacy infrastructure, regulatory barriers, low IT budgets and entrenched incentives won’t make AI adoption seamless - but they do create massive openings for those ready to build smarter, faster, and more patient-centric systems.

Healthcare won’t fix itself. The innovators who recognize this moment as an inflection point will define the next decade of medicine. I’m looking forward to seeing— and helping drive - this transformation.

Yet, recent developments suggest we are inching closer to a transformative moment for robotics. With breakthroughs in reinforcement learning, simulation, and multimodal AI, embodied AI - the fusion of AI with physical systems - is set to redefine automation. Following the panel discussion on “Embodied AI” at our AI House Davos 2025, I wanted to share some of the key advancements shaping the future of robotics. Are we on the verge of a robotics revolution, or are there still fundamental barriers to overcome?

Breakthroughs and Innovations in Embodied AI

The Role of Simulation in Robotics

One of the biggest obstacles in robotics has been the high cost and complexity of training AI-driven physical systems in the real world. Enter simulation. Today, advanced neural simulators allow AI models to learn and refine their behaviors before ever setting foot (or wheel) in the real world.

• Synthetic environments provide scalable training: AI systems can practice in high-fidelity virtual worlds, learning how to navigate, manipulate objects, and interact with dynamic environments.

• AI models can generate their own training data: With self-play techniques, simulators act as adversaries, constantly pushing AI to learn from novel and extreme scenarios that are difficult to collect from real-world data.

• Simulation bridges the gap between virtual and physical: The latest advances enable AI-trained in simulation to transfer to real-world settings with minimal adaptation, accelerating deployment timelines.

The increasing sophistication of simulation is a game-changer for industries like autonomous driving and industrial robotics.

AI-Powered Autonomy: From Trucks to Industrial Robots

Self-Driving Trucks Are Nearing a Commercial Reality

Autonomous vehicles have been in development for over a decade, but one of the biggest hurdles has been achieving full-scale deployment. While passenger car autonomy remains a work in progress, self-driving trucks are poised for a breakthrough.

Case Study: Waabi

• Waabi is set to launch driverless Class 8 trucks by the end of 2025, a major milestone for AI-driven logistics.

• The trucking industry has a severe labor shortage: With nearly 900,000 truck driver vacancies in the U.S. alone, automation offers a necessary and scalable solution.

• Simulation-driven safety testing: Waabi’s AI continuously trains in an advanced virtual environment, ensuring it can handle real-world unpredictability with unprecedented reliability.

Autonomous trucking represents one of the first commercially viable applications of embodied AI at scale, promising efficiency gains and enhanced road safety.

Industrial Robotics: Expanding Beyond Manufacturing

Robots in industrial environments are evolving beyond traditional fixed-arm automation. The latest generation of mobile robots is bringing autonomy to settings where human labor is scarce, such as inspections in oil and gas facilities, chemical plants, and warehouses.

Case Study: ANYbotics:

• ANYbotics’ quadruped robot, ANYmal, is already deployed globally to inspect and monitor industrial environments.

• Deep reinforcement learning enables adaptability: AI-powered robots can now handle stairs, dynamic obstacles, and real-world terrain with over 99% reliability.

• Embedded AI allows real-time decision-making: Unlike cloud-based AI models, these robots process sensor data locally, making instant adjustments based on their surroundings.

With industrial settings facing labor shortages and increasing safety regulations, robots like ANYmal offer a practical and scalable solution for physical-world automation.

The Challenges of Embodied AI

Why Robotics Is Harder Than Software-Only AI

Bringing AI into the physical world introduces new complexities beyond what we’ve seen with language models and software automation.

• Physical Constraints: Unlike AI that operates in the digital realm, robots must deal with physics: Weight, balance, power consumption, and durability all matter.

• Perception and Multimodal Inputs: Unlike software that processes text or images in isolation, embodied AI must integrate multiple sensor inputs: Cameras, LiDAR, radar, microphones—to accurately interpret the world.

• Safety and Regulation: Unlike software bugs, physical-world AI failures can lead to accidents. Proving safety before deployment is crucial, requiring robust testing and regulatory approval.

Humanoid Robots: Still a Long Road Ahead?

The hype around humanoid robots has grown rapidly, with major players investing heavily in this space. But are we truly close to humanoids performing useful work at scale?

Current humanoid robots excel in controlled environments, but they are not yet ready for complex, unstructured human interactions. Additionaly, Hardware limitations persist: Human-like dexterity, grip strength, and manipulation remain unsolved challenges. Practical applications will likely start in industrial settings before expanding to broader use cases.

So, while humanoid robots are making strides, we are still several years away from seeing them functionally integrated into everyday life.

The Future of Embodied AI: What Comes Next?

Safer, More Adaptive AI Systems

The future of embodied AI lies in creating systems that can reason, adapt, and interact more naturally with humans and environments.

• Contextual awareness will improve: Future AI-powered robots will better understand human movement, gestures, and speech, making them easier to deploy in collaborative settings.

• Hardware and AI will co-evolve: Smarter AI will reduce the need for expensive, high-precision hardware by compensating for minor mechanical inaccuracies.

• Regulatory advancements will shape adoption: Governments and industry stakeholders must work together to create safety standards that enable broader AI deployment while mitigating risk.

Final Thoughts

We are witnessing the dawn of embodied AI. While there are still many challenges ahead, industries like trucking, industrial automation, and logistics are already seeing tangible benefits from AI-powered robots. The dream of general-purpose humanoid robots may still be a decade away, but in many ways, embodied AI is already here—changing the way we work, move, and interact with technology.

As we continue to invest in this space at Merantix Capital, I’d love to hear your thoughts. Where do you see embodied AI making the biggest impact in the next few years?