Doctors are drowning in paperwork. Studies show they spend over a third of their workday (plus at least two hours per day outside of work) on clinical documentation, contributing to widespread burnout. But maybe it doesn’t have to be that way; a burgeoning industry developing AI medical documentation tools — or “ambient scribes” — and a growing number of physicians think this busy work should be delegated to AI.

The tools were originally designed to draft medical notes. Now companies are increasingly pushing them further. This includes providing context on the patient’s medical history, assigning billing codes and even acting on the next steps of a patient’s care.

“We can help not only with the notes, but we can hear your orders, so we can place your prescriptions or your diagnostic orders for X-rays or CTs. We can set the table for you on all these other tasks,” said Shiv Rao, founder and CEO of Abridge, a startup that in 2025 doubled its valuation to $5.3 billion within just four months.

Shiv Rao. Credits: Abridge

The category has grown quickly, from a few early-stage products to a market where startups, tech giants and healthcare companies all compete. Investors poured at least $1 billion into the space in 2025 alone, and Menlo Ventures in its 2025 State of AI in Healthcare report called ambient medical scribes “healthcare AI’s first breakout category.” (Menlo is invested in medical scribe company Heidi).

It has all the makings of a major boom, but the path to adoption isn’t anywhere near as simple as spinning up and funding products. The AMA’s 2026 Physician AI Survey found the adoption rate of the tools lags behind clinician enthusiasm for trying them — perhaps not surprisingly, placing AI recording software in between patients and their doctors comes with a lot of hurdles. While the tech and healthcare companies selling AI medical scribes have slightly different paths to market share available to them, they’re both up against rigid procurement processes and a whole lot of trust issues.   

Trust and process

According to Wiljeana Glover, a researcher focused on healthcare innovation at Babson College, the gap between adoption and interest is mainly down to the procurement process itself. Bringing on such a technology is a big decision in these settings, and medical groups have very formalized processes to identify their needs, select potential candidates, screen the top choices and evaluate them on specific criteria. “That process can take time,” she said. 

Wiljeana Glover. Credits: private

Early studies suggest the tools are effective at reducing clinician time spent on documentation, but they’ve been limited to one healthcare system or have been conducted over short periods of time. Jamie Chang, chief medical officer at WellSky, a healthtech company with an AI medical scribe product, said the challenge is getting the decision makers to trust the tool will be accurate and won’t create any additional burdens.

“I think the value prop and desire is there, but maybe [it takes] more trust in the process,” he said. 

To sell to big medical groups, Rao described having to go through several executives including the chief medical officer (CMO), CFO and CIO. The CMO wants to know specifically what problems you can solve for their clinicians, while the CFO is interested in the bottom line, he said. The CIO needs to be assured you meet all the privacy and security requirements, and will be able to serve them for years to come. At the same time, selling to providers outside those big health systems comes with its own challenges, in particular just trying to get in front of them. 

Jamie Chang. Credits: WellSky

“The go-to-market for those clinics can be very painful,” he said. “It's hard to knock on all those doors and be able to scale something quickly. So finding channels for distribution becomes really key, especially for startups who are fighting incumbents or hegemons.” 

With companies of starkly different sizes from both the healthcare and tech industries offering these products, Glover said she’s seen their ability to effectively go through the procurement process vary. She’s also noticed a pattern of the large health systems choosing for consistency, even if it means sacrificing on customizability of the tool.

WellSky, for example, is now giving away its AI medical scribe offering for free to customers already using its platform for other functions, which Chang said is “absolutely” an advantage as options flood the market.

“I think over time, the industry continues to consolidate rather than have a lot of bolted-on options,” he said. 

Patient perspective

Ambient medical scribes have potential benefits for patients, too: a doctor who isn’t heads-down taking notes during your appointment or generally burned out can theoretically provide better care. Yet this potential upside isn’t obvious to patients and doesn’t necessarily outweigh concerns about AI recording their sensitive medical conversations, especially as overall distrust of AI increases.

In one survey of patient attitudes toward ambient medical scribes conducted at Stanford Health Care in 2025, 73.6% of 2,202 patients who experienced their clinician using it said they would like it to be used again. They were told in advance that the visit would be recorded and given the chance to decline. Without first-hand experience and upfront disclosure, however, patient willingness looks different. A much larger survey of 12,153 Canadian adults based only on perception and not a direct experience found that only 28.3% were aware clinicians were using AI scribes at all, and that, even while anticipating benefits from its use, 61.8% were reluctant to have an AI scribe used in the future. The researchers cited privacy concerns and low awareness as key barriers.

“The ethical risks fall disproportionately on patients... [Ambient AI systems] generate vast amounts of data that may have secondary uses — whether anticipated or not — raising significant concerns about privacy, surveillance and data governance,” wrote the authors of a 2026 comment on consent and ambient clinical AI.

The authors also argue current deployment practices fall short of getting meaningful patient consent, including not adequately explaining recording practices or providing opportunities to opt out. Several healthcare providers have already been sued by patients in class action lawsuits claiming the technology was used without consent.

On patient privacy, Rao said a lot of Americans now “just expect that things are getting captured in some way.” and that AI meeting note-takers have normalized the idea. Doctors still need to get consent and spend time talking to patients about it, he said. Chang noted physicians he’s worked with are facing patient questions about how the recordings are used. This is putting doctors on the spot to talk about AI and data practices they might not even fully understand themselves, said Glover, adding that the tools should be designed with privacy features top-of-mind, such as automatically filtering out sensitive data.  

“Keeping patient needs front and center allows health systems to truly make wiser decisions,” she said. “The influx of a lot of shiny objects, as we might call them, can become really distracting.”

Oumi started out as a popular open-source project. This meant it had interest and validation before it even became a commercial business. Something that enabled the company — whose platform lets customers automate building purpose-specific AI models — to quickly garner a bit of income from small startups that jumped on the platform after it officially launched. But for co-founder and CEO Manos Koukoumidis, Oumi’s first deal with a major enterprise was the true “first customer” moment he felt really validated his business. 

“[Startups] are easier to convince. It’s easier to change the strategy. But when such a large organization [signs on], I think this is much more telling than getting a small startup,” he said. 

For Oumi, that first large enterprise contract was with a top-10 global bank the company declined to name. The deal came about after Oumi presented at an in-person conference hosted by its investors. The pitch landed for the bank right then at the event, but it took about five meetings with the bank's senior leadership and its engineering leads to actually close the deal. Oumi said the bank was drawn to the combination of automated model development and ownership of the resulting model. 

Koukoumidis and his co-founders all come from Big Tech, previously working at Google, Meta and Microsoft. So on a personal level, he said getting that validation of a product and idea that was fully his own was really rewarding.

“Building your own company on a strong conviction that you have, especially when you start on something that's so contrarian and you hear people who are industry experts, most reputable VCs, saying ‘you're doing it wrong.’ And you’re like ‘no, wait and see. The time will come,’ and then you see that shift happening. You see the biggest organization in the world sitting and willing to sign a contract with you,” he said. “It’s very rewarding to say the least.”

That shift is bigger than one bank. Read the full article on DeepL, Loora and Oumi, and the third option opening up between building an AI model and buying one.

Top photo: Manos Koukoumidis. Credits: Oumi

Where did you get your AI model? In 2024, 53% of enterprise AI solutions were purchased rather than built internally, according to Menlo Ventures. A year later, that share rose to 76%. 

Nick Patience, vice president and practice lead for AI at The Futurum Group, attributes that uptick in the “buy” strategy to two factors. A lot of 2024-era build efforts didn’t go so well (often due to a lack of data readiness), while the frontier API market matured, offering lower upfront cost and faster time to value. 

But two things changed in 2026. As AI providers move to usage-based pricing models, customers are now grappling with unpredictable and often exorbitant inference bills. Secondly,  access to the models themselves turned out to be conditional. When Anthropic temporarily suspended access to its Fable 5 and Mythos 5 models to comply with U.S. export controls, it made the risks of relying on models you don’t control very real.

Nick Patience. Credits: The Futurum Group

Some companies that have been building their own models argue it’s allowed them to save costs, achieve better quality, and maintain control. But training a model from the ground up is still a major endeavor. For companies that want all of the benefits without committing to a full build, a third option is emerging.

“Training a purpose-specific model on open weights that the company fully owns and can run anywhere, that's a distinct third category from either buying an API or building from scratch — and it's the one gaining the most real traction this year,” said Patience. “It gives portability without the enormous cost of pretraining, which is why I'd expect it to keep growing faster than either pole.”

Building from scratch

DeepL is one rare example of a company that chose to train its own translation-specific model.

Launched in 2017 and long a leader in translation, the company’s product predates the LLM boom and was built on earlier neural-network technology. But when DeepL decided to redesign its product with a new LLM-based architecture in 2024, the decision to build was an easy one for co-founder and CEO Jarek Kutylowski.

“[We’re using] LLMs that are really specialized for the kind of translation task that we're doing, and it's extremely important when you want to maintain certain quality levels and create models that are actually excelling at this particular task. You don’t want the model to be doing everything,” he said.

Loora team. Credits: Loora

Roy Mor, co-founder and CEO of English-learning company Loora, said that building a language model from scratch these days “is a crazy thing to do,” arguing there’s no reason to go that route because you don’t gain anything. But when it came to integrating a new video model to power the mobile app’s AI avatar, developing a custom model in-house won out on both cost and quality.   

“The cost [of the off-the-shelf solution] just doesn't work out from a unit economics perspective. It's like 10x [the cost] for ElevenLabs, and also they're really not built for on-device and to be quick,” he said. 

Because the scope was narrow, Mor said they were able to train a really small model using a custom architecture that runs on-device, eliminating the network round trip and creating a better user experience at a much lower cost.

This is where Patience says building makes more sense: when an AI capability is part of the actual product, pricing or competitive differentiation, or when the workload has ownership economics beating ongoing usage-based fees over a multi-year horizon. Another reason to build is when regulatory requirements make dependency on a single external provider a business risk rather than a convenience, he said. 

A new way to own your model

Oumi, a Seattle-based startup founded by former Google and Microsoft AI engineers, is betting more and more companies will want ownership over their models, but they won’t want the burden of building them. 

As an answer to this, Oumi created a platform to enable customers to automate the post-training of open-weight models to build AI models for specific use cases. Users start by describing the model they want to build in a prompt box using natural language. The platform then helps them compare open-weight models and curate their data before building it out. Models built with the platform are infrastructure-agnostic and can be exported and run wherever the user wants. 

Manos Koukoumidis. Credits: Oumi

“You own everything. You own the model weights, the data, the recipes, which is how every step was executed,” said co-founder and CEO Manos Koukoumidis, adding that he believes companies need optionality and to not be beholden to anyone else for this vital technology.  

Koukoumidis said their customers are seeking quality differentiation and control. When AI is a big part of their product, they feel they’re not differentiating if they’re using the same AI as everyone else, he said. Others want full control over privacy, security, where they deploy models and to make sure no one can take them away, he said. Because of the latter in particular, he’s found Oumi has especially resonated with companies in regulated industries. Oumi’s first enterprise customer was a top 10 global bank, which he said wanted higher quality, lower costs, and full control. 

“I think the mindset is shifting,” said Koukoumidis.

Top photo: Loora co-founders Roy Mor and Yonti Levin Credits: Loora

Foundation models in biology have largely focused on drug discovery. London-based twig bio is applying the same approach to biomanufacturing, using its model Canopy to design microbial strains that produce ingredients for consumer goods like sunscreen, cosmetics and personal care.

Canopy is trained on more than three years of proprietary experimental data generated in twig’s lab, and is designed to shorten the path from metabolic engineering strategy to a production-ready strain. Its first product, Mycolux, is a fermented alternative to petrochemical-derived ingredients used in sun protection.

Satnam Surae, twig's chief technology officer, spoke with The Infinite Loop about  how proprietary experimental data changes strain design and what worries him most about getting a working microbe out of the lab.

'We built the data first, then the model'

TIL: A company comes to you and says, "We need a fermented substitute for material X." What do you do next?

Surae: What we do, and how we help customers, is replace chemicals used predominantly in personal care and cosmetics. We do that by leveraging Canopy, our multimodal knowledge-graph foundation model, which helps us design metabolic engineering strategies: the microbes we then build in our lab and test in fermentation systems to produce those molecules at scale. Canopy lets us do that quickly, effectively, and with high confidence that when we get to the expensive part of the process in the lab, those designs will actually work.

TIL: Most people, when they hear “foundation model” in biology, think of drug discovery. Why did you go toward bioengineering instead?

Surae: I've worked in pharma, but I've also worked in the chemicals industry, and there's a real gap in terms of specialized models for it. We lean on more generalized protein models and take inspiration from a lot of the pharma-focused ones, but we needed to build this model ourselves because nothing out there would really help us. We took the initiative: we saw the gap, we understood the space, and we've been building data in our lab for more than three years to get here. We built the data first, then the model, and now we can iterate quickly and generate more data to accelerate the model's capability.

'No one had ever taken these molecules to market'

TIL: Who are your customers?

Surae: We're selling to specialty chemicals customers and consumer packaged goods companies. We currently have three customers: one in the US, one in the UK and Europe, and one in Asia, so we're already spreading geographically. Our first product comes to market later this year — it’s a novel biomolecule to replace the chemicals in sunscreen. That's a massive opportunity, because most sunscreen chemicals are endocrine disruptors. They're environmentally unfriendly, and critically, they come from petrochemical sources, which means they're exposed to supply-chain and price volatility on top of the sustainability problem.

TIL: How did the idea of this product come about, and what did it take to get to market?

Surae: Mycolux is a formulated ingredient that goes into end products. We sell it to customers, who package it into the formulations consumers actually buy. The idea came from a gap we spotted, a real need for innovation in sun protection, with few good options on the market. We understood the science and knew it was feasible, but no one had ever taken these molecules to market, so that was the risk. The traction's been fantastic: going from nothing to a production strain in manufacturing-like conditions took under 18 months, which for us is actually slow. With Canopy and the acceleration it gives us, we're confident the next one will take half that, if not less.

Satnam Surae and twig’s CEO Russ Tucker at Nebius AI Discovery Awards. Credits: twig

TIL: How far does Canopy's scope extend: is it just sunscreen and everyday chemicals, or more than that?

Surae: Much more. We built Canopy to be as multi-purpose and generalizable as possible for biomanufacturing. Beyond the sunscreen molecules, there are related ones that work as antioxidants, others that could go into flavors and fragrances, and others still that could be pharmaceutical ingredients, even drug molecules. Really, across the board.

'It's more engineering than science'

TIL: What's the failure mode you're most worried about?

Surae: It's more engineering than science. We can prove the science works in the lab; the real risk is getting something that works in the lab to work at 100,000-liter scale as an engineering process. Iteration is slower there, costs are larger, and it comes down to techno-economics more than anything technical or scientific. That's part of why Canopy factors in biomanufacturing-relevant data when it runs inference: it minimizes that risk as much as we can. Our mantra has always been to fail quickly and generate data even from failed experiments, because that data is still valuable.

TIL: And on compute: is this an inference-heavy workload?

Surae: There are two parts. Training is heavy, but critically, so is inference: we're running thousands of separate analyses per molecule, across thousands of molecules, in parallel. That gets extremely heavy. But partnering with someone like Nebius on inference, going forward, is something we're looking at to help maximize that potential.

TIL: You picked up second place in the Genomics category at the Nebius AI Discovery Awards in July 2026. What's your plan with the credits?

Surae: We’ll use the credits to accelerate our scaling plans for Canopy. We’re increasingly training across hundreds of millions of biological and experimental data points, so access to large-scale GPU infrastructure with the requisite availability is critical.

Disclosure

Surae spoke with The Infinite Loop at the Nebius AI Discovery Awards ceremony in London, where twig bio won second place in the Genomics category. The Infinite Loop is a Nebius editorial project. Editorial decisions are made independently.

Arkangel AI is a Bogotá-based company that analyzes medical data to flag patients who may have a missed diagnosis. It reads the parts of a record that many healthcare AI tools skip: physician notes, shorthand, images and other unstructured material that makes up much of what a clinician writes down.

Founded in 2019, Arkangel now operates across Latin America. Its clients include Novartis, AstraZeneca and Roche.

Arkangel's co-founder and president Laura Velásquez Herrera spoke with The Infinite Loop in London about the customer that arrived before the product was ready, and the investor who told her she was nothing.

TIL: You sell to both pharma and hospitals, buyers with very different incentives. What’s the connective tissue?

Velásquez Herrera: Arkangel is an AI company. What we do is analyze medical records to identify different types of diseases early enough. So we can help patients, we can help the physician, and we can help the hospital and the healthcare system.

We work with three different types of customers. Healthcare insurers, where we analyze very large datasets to identify errors on both the operational and the clinical side. Pharmaceutical companies, where we work with AstraZeneca, Roche, Sanofi, Boehringer and so on. And hospitals, where we help them identify those overlooked patients.

So we help hospitals find patients, we help the pharmaceutical companies increase access to different types of treatment, and we help the healthcare insurers analyze all this data in order to help the patient as well. It is basically the same technology, but the output for each of them is completely different.

TIL: If people ask what your moat is, what makes you different, what do you say?

Velásquez Herrera: What we have been working on is a way to analyze unstructured data to identify the patients who would otherwise be missed. What you can find in the market right now are specific algorithms that analyze structured data in order to give you a specific output for a specific disease. What we are doing is analyzing any type of data. It could be images, it could be any type of notes from physicians. We can connect the data points to identify issues and reach patients with an early diagnosis. And we have scientific publications that validate what we have been working on, including a paper presented at the 2025 IEEE Conference on Artificial Intelligence on PANDORA, our framework for extracting clinical data from unstructured records and applying clinical risk scores.

Co-founder and president of Arkangel AI Laura Velásquez Herrera. Credits: Nebius

TIL: What sets the Latin American market apart?

Velásquez Herrera: I am from Colombia. We do not have much access to healthcare in Latin America. Right now, we are trying to close that gap. We are operating in Mexico, Colombia, Peru, Argentina and Chile, helping patients through the healthcare system.

Where it started

TIL: How did you start? What was the initial idea?

Velásquez Herrera: We have been working on it for about six years. We started because of a family problem. I lost three of the most important people in my life, in my family, just because they were not diagnosed at the right time. And I lost all of them in less than a month and a half.

And I said, how can I help my family take a more preventive approach? Then we started building a chatbot on Facebook Messenger, and it went viral. In less than a month, we had more than 9,000 users in 22 countries. And we said, okay, this is a big problem. We need to understand what is going on here. That is why we started.

TIL: What was your biggest obstacle in building Arkangel?

Velásquez Herrera: The biggest obstacle, I think, was that I did not know anything about health or AI, but I started building this AI startup in the healthcare space. So building trust was the most difficult thing.

We were living in Canada and the U.S., and when you start building this type of thing, being a woman and being Latina, people say, okay, you don't know anything about this. You have to be more conscious about what you are doing, and you have to put in more effort to make it possible. So both of those things were hard at the beginning.

But right now everything is different when you can prove that it works. Even if I am a woman or Latina, we can build something big if we can help people.

TIL: When did it shift from hard to possible?

Velásquez Herrera: I think nothing changed. Maybe it was me. I changed my mindset about being able to do it even if other people didn't believe in me. I was in Montreal, Canada, pitching at a demo day. And an investor told me: "You are nothing because you are a woman and you are Latina, and this is nothing. Next."

And I thought, oh my God, what is going on here? I started thinking about it, and I said, okay, the problem is not mine. The problem is his, because he is trying to push us back just because we are not from the same background. And this changed my mindset.

The world can shit on us, sorry for the word. But it is your mind, it is your problem to change your behavior and your attitude when approaching these kinds of things. Of course, sometimes it is hard, sometimes it is easier, but it depends on how you approach things.

‘You feel like you are doing something that matters’

TIL: We run a newsletter format called The First Customer, where we ask founders about the moment a company got its first paying customer. Who was yours?

Velásquez Herrera: It was Novartis. They were looking for something new to identify patients with malaria early enough in rural locations in Latin America. We said, okay, we can apply for this, let's see what happens. But we did not meet all the requirements they were looking for.

And suddenly they said, we want to understand a little bit more about what you have been working on. We were doing everything from scratch. And they said, okay, we really like what you are doing, we are not sure what is going to happen with this, but we want to try. How much does it cost?

And we said, okay, we have to get the hospitals, we have to get the patients, we have to get all of these things and prove that it works in rural locations, with no internet, no anything. And we agreed to try.

We implemented it in Chocó, a rural region in Colombia. And we helped almost 6,000 people by identifying two diseases: malaria and Chagas. Actually, we started building AI for rural locations, not for the cities.

TIL: How did it work technically?

Velásquez Herrera: We were reading images of the blood smear, what you see through the microscope. We were analyzing what type of malaria a person had, and the parasitemia, how many parasites they had. We analyzed those things by taking a picture with a phone used by people there. It wasn't an iPhone, it wasn't an Android, it was the worst mobile anywhere.

So we had a lot of barriers, but then we proved that if we can bring AI to rural locations, we can help a lot of people for free. So we started building this business model in which we ask pharma companies to pay for it, and we implement it in different locations. It can also work in urban locations, because that is where most people live, but with this we can help people in rural locations for free.

When you can help people, it is really nice. You feel like you are doing something that matters. The people we help will never hear the word ‘algorithm.’ They may never know Arkangel exists. They just get diagnosed in time. That’s the only metric that really matters to me.

TIL: You've been awarded the regional trailblazer award at the Nebius AI Discovery Awards. What’s your plan with the credits?

Velásquez Herrera: We plan to use the credits to adapt Pandora to new hospitals' documentation styles and make it faster and cheaper to run as we scale.

Disclosure

Velásquez Herrera spoke with The Infinite Loop at the Nebius AI Discovery Awards ceremony in London, where Arkangel AI was named a regional trailblazer for Latin America. The Infinite Loop is a Nebius editorial project. Editorial decisions are made independently.

When Zach Rash and Brad Squicciarini founded Coco Robotics in early 2020, the timing could hardly have been better. Or worse.

The Los Angeles startup launched just as the COVID-19 pandemic arrived, accelerating demand for doorstep food and grocery delivery almost overnight.

The timing created both an opportunity and a challenge. While demand for delivery surged during the pandemic, supply chain disruptions made it difficult for the young company to source reliable components. Coco's early robots were built using mostly off-the-shelf hardware and, according to Rash, "the product was awful."

Rather than spending years on research and development, Rash and his team took a different approach. They started making deliveries almost immediately.

Coco's first customers were L&K Market, a corner store in Santa Monica, and Alfalfa, a restaurant that had recently opened in the neighborhood.

Rash found them the old-fashioned way: by knocking on doors.

"We want to build this with you," he told them.

The pitch was far from risk-free. Restaurant owners care whether orders arrive on time and customers stay happy. Introducing an unproven delivery robot into a busy operation could easily have created more problems than it solved.

Instead, the owners embraced the experiment.

"They loved the opportunity to help us shape it," Rash said.

Their feedback became a blueprint for the company. Rather than building features that engineers found interesting, Coco learned directly from the businesses that were using its robots every day.

Today, Coco operates more than 1,000 delivery robots and works with partners including Uber Eats, DoorDash, Taco Bell and Shake Shack.

For Rash, the lesson was simple: do not build technology in isolation. Build it with the customer standing beside you.

Avride delivery robot at Indiana University. Credits: Avride

Avride: Proving it in Austin

Around Avride's engineering center, the team set up obstacle courses to demonstrate how the robots handled different situations. Uber's team was also shown the machines "from inside out," including how their hardware and autonomous systems were designed for safe operation. Then the demonstrations moved onto public streets.

"The real-world demonstration was the most important," said Yulia Shveyko, head of public affairs at Avride. "Everyone wants to see how our robots operate in this unstructured sidewalk environment."

The demonstrations were part of the process that led to Avride launching its delivery partnership with Uber Eats in November 2025, the Avride’s first major commercial relationship.

"Prior to the actual launch, we were preparing an actual safety plan for Uber to approve," said Roman Nefedov, head of delivery robots at Avride.

Safety was central, but so was efficiency.

“Robots should not only be safe, but they should be efficient,” Shveyko said. “Any food tech company doesn’t want to get into a situation where a robot would take an hour to bring food from the restaurant to a customer.”

Delivering via Uber Eats app. Credits: Avride

Asked what ultimately convinced Uber that Avride was ready, Shveyko pointed not to a single breakthrough, but to having met the platform's requirements.

Shveyko’s advice to other robotics startups looking for their first major customer is simple: "Be honest."

Read the full story about how Coco Robotics and Avride are teaching their robots to survive a city that refuses to behave.

Avride is part of the Nebius group, which owns The Infinite Loop - Ed. 

If you've ordered takeout from Uber Eats in Austin or Miami, there's a chance your food was delivered by a robot. What looks like a last-mile delivery is also a training exercise for the next generation of autonomous systems.

Companies including Coco Robotics and Avride are using commercial deployments to generate the operational data needed to improve autonomous systems and expand delivery into more cities.

What cities teach robots

For Zach Rash, co-founder and CEO of Coco Robotics, every commercial deployment contributes to the dataset used to improve future AI models.

"We have millions of hours of driving in these sorts of environments," Rash said. "That can be used to train a really reliable self-driving model on the robots."

City streets are simply too unpredictable to reduce to a fixed set of rules.

Coco’s delivery robot on its way. Credits: Coco Robotics

"The reality is you're just always surrounded by people on e-bikes flying by you and cars and people walking and people coming in and out of stores," he said. "Creating those rules is impossible."

When robots encounter unfamiliar situations or become stuck, operators can step in to help: that human intervention becomes data too.

"The best way to create that dataset is to provide actual value, economic value, so you can continue to collect that data and iterate from there," Rash said.

The biggest difference from self-driving cars

Avride has reached a similar conclusion from a different starting point. Rather than building delivery robots first, the company adapted technology originally developed for autonomous vehicles.

"The overall goal for us internally is to reuse as much as possible from the bigger brother, from the car project," said Roman Nefedov, head of delivery robots at Avride.

Much of that autonomy stack transferred intact. The robots share localization, perception and some prediction technology with Avride's self-driving cars, allowing them to identify objects and anticipate how the world around them is likely to evolve. 

The biggest difference lies in planning. According to Avride, cars operate in a structured environment of roads and lanes, where other road users are expected to follow a set of rules. Sidewalks, by contrast, are a completely unstructured environment of pedestrians, cyclists, dogs, street furniture, and temporary obstacles. 

Delivery robots on a mission. Credits: Avride

"The algorithms for feeling the world are the same," Nefedov said. "The algorithms for planning what to do next, they're completely different."

Understanding the world also means knowing precisely where the robot is within it. GPS can identify the correct street, but often lacks the precision needed to locate the exact front door, apartment entrance, or collection point.

To address that challenge, Coco partnered with Niantic Spatial, whose Visual Positioning System combines geospatial AI and computer vision to localize robots when GPS alone lacks sufficient precision.

"The maps that you need for robots are quite different from those for a person," Rash said. "The good thing about robots is you can tell us exactly where to go, and we will go exactly there. The bad news is you might not have picked the right spot."

Engineering for the real world

Operating hundreds of robots every day requires different engineering. One challenge is deciding what intelligence belongs on the robot and what can be handled elsewhere.

Human oversight remains an important part of that process. Yulia Shveyko, head of public affairs at Avride, said the company's onboard perception and prediction systems can recognize individual objects, but have a more limited understanding of the overall scene.

To add that context, Avride uses a cloud-based vision-language model (VLM) that receives camera snapshots from robots every few seconds and flags situations that may need human monitoring. The VLM does not control the robot, avoiding the latency and connectivity risks of cloud-based control.

"The VLM itself does not control the robot; it does not send any commands," Shveyko said. "It works as an alert system."

That can help determine whether a first responder is actively dealing with an emergency, for example, or identify wet cement at roadworks rather than simply recognizing cones and barriers. When the system flags a potentially difficult situation, a remote assistant can monitor it and, if necessary, send instructions for the robot to execute locally. 

As fleets grow, managing the data they generate becomes an engineering challenge. Every robot continuously records video, lidar, telemetry and other sensor data, producing up to a terabyte of information during a day's operation.

"The engineering team always wants more logs because data is just the fuel for improvement," Nefedov said.

Uploading all that data is expensive. Rather than transferring everything over mobile networks, Avride selects the most valuable moments for future model development. Those decisions matter more as commercial fleets expand.

Not SOS, just ZZZ  

Energy efficiency creates another set of trade-offs. Autonomous delivery robots consume roughly the same amount of power as a walking human, but around half of that energy goes to sensors and onboard computing. Every watt spent processing AI models is a watt unavailable for extending operating range.

Coco delivery robot in the snow. Credits: Coco Robotics

For Nefedov, those engineering compromises distinguish commercial deployments from technology demonstrations.

"Energy is a huge concern for us, and that's what differentiates a demo from real work."

That focus extends to surprisingly small details. During quieter periods between lunch and dinner, robots shut down as many systems as possible to conserve battery power. Even the LED display was redesigned to minimize energy consumption after pedestrians mistook stationary robots for broken machines. Instead of switching the display off completely, the robots simply show "ZZZ" to indicate that they are sleeping while waiting for the next delivery. Shveyko said reports from people concerned that stationary robots were stuck have since dropped dramatically.

Learning in the real world

For Coco, expanding beyond Los Angeles required more than retraining AI models. 

"When we met with the Finnish regulators, their comment to us was that these winters took down the Russian army," Rash recalled. "Are you sure you want to launch here?"

Operating in those conditions required heated cameras, weather-resistant hardware, and new operational data for future deployments.

Avride found that seemingly simpler environments could prove equally challenging. Nefedov said university campuses, despite having fewer vehicles, often proved just as challenging because dense pedestrian traffic demanded a different style of planning.

Coco and Avride have adopted different technical approaches but describe the same underlying dynamic: commercial deployments generate operational data, operational data improves AI models, and better models enable broader commercial deployments.

Finnish winters, crowded campuses, a dog that wanders into frame: none of it can be simulated away. The city still refuses to behave, but it's the only real teacher either company has.

Top photo: Delivery robot on the streets of Austin, USA. Credits: Avride

Avride is part of the Nebius group, which owns The Infinite Loop - Ed. 

“Why do I need to die? I really enjoy experiencing life,” Filippo Nigro said when asked why he launched his longevity startup You(th).

“It’s actually potentially possible to reverse aging. These are solvable problems according to the law of physics. When I discovered that I said, ‘Okay, even if there is 0.1% probability of reaching longevity escape velocity during my lifetime, then I'm just going to put blood, sweat and tears into trying to solve that.’”

Longevity escape velocity refers to a future when humans can use science and technology to extend their lifespans faster than they age, meaning that, theoretically, they never need to die. Nigro identifies two factors required to make this a reality: biotech breakthroughs in gene reprogramming, and a scalable platform for measuring people’s health that can predict potential risks before they occur.

It’s the second piece of the puzzle that he’s working on with his company You(th), which combines smartphone cameras and microphones with clinically validated AI models to let users screen for more than 50 biomarkers.

Nigro’s mission to give people a better understanding of their overall well-being is part of a broader trend. Specialized “longevity clinics,” which offer advanced health screenings for patients, have quadrupled in number since 2020.

Now startups like You(th) are trying to make the benefits of these elite and expensive services more accessible, using AI and information from smartphones and wearables to let people better track their well-being. But with a potentially limitless stream of data to work with, there are questions about how to build the infrastructure to support these products as they scale, not to mention whether playing god in the pursuit of living forever is really a good idea.

“Making sci-fi a reality”

Nigro co-founded You(th) in 2023, after coming across research showing that smartphone cameras and microphones could be used to detect biomarkers for conditions ranging from cardiovascular well-being to prediabetes and mental health. The app licenses clinically validated, third-party AI models that allow users to check biomarkers such as their blood pressure and heart rate variability, based on short selfie videos, photos and voice recordings.

“The camera measures how light is reflected on your forehead, on your chin, and so on. And the light is different depending on blood flow, so you can detect: how fast does it flow, how strong does it flow, how often does it flow,” said You(th) co-founder, Johannes Kraus. 

A product shot outlining some of the biomarkers You(th) can test for. Credit: You(th)

The ability to test for these kinds of biomarkers with such widespread consumer technology might be impressive, but You(th) now has a challenge in educating the market that its product actually works.

“The biggest barrier to product-market fit is trust, because you are making sci-fi reality,” said Nigro.

You(th) co-founder Filippo Nigro. Credit: You(th)

After initially launching as a customer-facing app, You(th) now runs a B2B2C model, partnering with clinics and insurance providers looking for ways to keep their customers more engaged with their own well-being. Beyond getting the You(th) app into more users’ hands, Nigro said that working with established health providers is a helpful way to build confidence in AI’s capabilities: “We’re scaling B2B2C to build trust.”

“Homo Superior”

Another longevity startup using AI to help people understand their own well-being is Dubai-based AVARA, which combines data from wearables with blood test information to deliver personalized nutrition protocols for its users. 

“The one thing we do from the moment we're born to the moment we die is eat food, that's a guaranteed nonstop. Why is that not personalized?” said co-founder Josh Sharron. 

AVARA’s platform draws on a dataset of more than 1,000 peer-reviewed sources from research into how different foods affect different biomarkers, to give users insight into how their diet influences their bodies.

“We hook up the wearable, then we get the data from sleep, your recovery, what your actual habits look like, then we're able to hyperpersonalize your meals, because we know what you're lacking in which biomarkers,” Sharron said.

The AVARA app, showing nutrition-biomarker optimization. Credit: AVARA

He believes that, by giving people a better understanding of how nutrition affects their well-being, AI can help move humanity into a new, sci-fi-inspired era, where people can relentlessly optimize themselves in pursuit of longer and healthier lives.

“In X-Men they call [the mutants] ‘Homo Superior.' You’ve got Neanderthals, then Homo sapiens, then the next stage of evolution is [like a] Homo Superior, where we’ve got so much tech and data that we can optimize ourselves in every way that we're actually pushing the frontier of human evolution,” said Sharron.

Cheating death?

But moving humanity to its next stage of evolution could come with a big token bill. You(th)’s Nigro said that, in the future, users won’t even need to open the app to give it the information it needs to assess their biomarkers, as the company will be able to use “passive” data, gathered from their wider digital footprint.

“You’ll just unlock your phone, take stories on Instagram, send voice messages on WhatsApp, right now I'm speaking with you [on Zoom], generating facial and voice data,” he told The Infinite Loop. “I can basically have a longitudinal assessment about my blood pressure, metabolic health, respiratory health, cognitive health and skin health from that data.”

To make processing all of this data feasible, startups like You(th) are trying to move as much processing as they can to the edge to avoid hefty cloud-inference costs.

“You will need to have quite a sophisticated system. For example, our facial video data is processed on-device; it doesn’t go to the cloud,” said Nigro. “We just need to watch out that the app is not so loaded by adding so many models onto the device that it then takes ages to download the app.”

When asked whether he believes critics of the longevity movement who say that aging is a natural part of life that shouldn’t be pathologized, have a fair point, Nigro responded saying that people will get with the program.

“Society needs to adapt. Right now, it might feel unnatural, but in the early 1900s we were not even flying, so planes were unnatural because people couldn’t fly, and 60 years later we’re flying to the moon,” he said.

But, while Nigro’s motivations to cheat death are on the more extreme end of the scale, AI platforms like You(th) and AVARA might also help those of us who just want to stay healthier for longer. At a time when public health care systems are under increasing strain, the move to a more preventive system that spots potential problems before they arise, rather than treating them reactively, might be no bad thing.

Thomas Baucom’s calendar perpetually looks like “a bad game of Tetris.” Working as a forward deployed engineer (FDE) at productivity platform Superhuman, he’s often on back-to-back calls talking to customers across industries, hearing out their problems and requests. Sometimes he’s jumping on a plane to be by their side.  

Previously a customer success manager, he stepped into his new role two months ago. Baucom said the job is similar in that it’s also focused on customer insights. As a customer success manager, he was also talking all day, every day with customers about how they wanted to use the products. The key difference is now he can spring into action on what he hears.  

“In that role before, I had to say, ‘well, we're working on that,’ or ‘we can put that in the backlog, but it doesn't work like that today,’” he said. “The big difference in the FDE role is I can now say, ‘I can fix that.’ I have the ability to work in our codebase, go on-site with you and build those things out. I can go much much deeper.”

That more direct line from insight to action is why companies of all stages and sizes are embracing the FDE model, which revolves around embedding directly into customer organizations to help them adopt the product, solve specific problems and scale usage. 

Job listings for FDE roles increased more than 729% between April 2025 and April 2026, according to data from Indeed shared with Business Insider. Amazon in June announced a $1 billion investment toward a new FDE unit, and just two days later, Microsoft announced its own $2.5 billion FDE initiative. Anthropic, OpenAI and Salesforce have also all announced similar initiatives earlier in 2026. 

Now that some adopters are well into their FDE journey, they’re starting to see the outcomes — in particular, rapid, iterative speed to deployment. They’re also learning some lessons around how to leverage the FDEs effectively.

Need for speed

Superhuman started embracing the FDE model around four months ago, and Baucom said he’s already shipped around 20 customer projects. He’s seen first-hand that speed is the most undeniable benefit, and others agree.

“What was taking months is taking weeks,” said Mike Park, CIO at Infor, which has done a big push into FDE engineering starting about a year ago. 

Kevin Wu, co-founder and CEO at Leaping AI, a voice AI startup that hired its first FDE last year, also cited speed as the biggest benefit and said this is particularly true when they go on-site. 

Kevin Wu with two recent FDE hires. Credits: Leaping AI

“Flying to the customer actually helps because you get everybody in one building, and then you just finalize the last few steps on day one,” he said. On day two, they train the customer on how to use it and monitor the performance to see what can be immediately improved. After that, they gather feedback, integrate it, and have weekly virtual check-ins from there on out. 

Beyond speed, FDE teams have noticed that customers appreciate the hands-on approach.

For one example, an Infor FDE worked with a customer in industrial manufacturing to build a system that flags itself when it needs attention, so the customer no longer has to watch the dashboards constantly. The challenge with a use case like this, Park explained, is that companies have different levels of comfort trusting AI. The FDE worked with them directly to understand their tolerance and tailor the solution to their exact stage of AI adoption, and help them feel more comfortable along the way, he said. 

Mike Park from Infor. Credits: Infor

For a recent project, Baucom transformed a multinational publisher’s 400-page style guide into a series of custom agents, with one for each of its many sub brands. As a writer works, the agent reads the text and flags any deviations from the relevant style guide. It underlines errors for the writer to review, shows the correction and points to the corresponding entry in the style guide. The customer got a fast turnaround and a bespoke solution built around keeping its brand consistent.

“We're going to cater this solution to you. We're going to make it easy. We're still going to have the human element there, and that seems to really resonate at least with folks I get on the phone with.”

The part that isn't automated

Overall, the main lesson Baucom has learned so far about being an FDE is that AI is still intimidating for a lot of folks, which makes a strong human touch more valuable, not less.

“People still want to have a Thomas they can Slack,” he said.  

While FDE is typically thought of as a customer-facing role, companies have also learned the benefits of the model can be leveraged internally. Baucom said he’s built prototypes for tools for his teammates to be able to run better demos and agents to help prep for calls, for example. 

For Park, another lesson from Infor’s FDE work so far is how crucial it is to stay up to speed. As models progress, teams need to keep evolving and iterating quickly, he said. For this reason, he believes the fast pace of AI itself is a driver for the FDE model.

“With this ability to iterate faster, I think [the FDE model] allows you to take advantage of the technology even faster than you would have before because of this concept,” he said. “And I think that's where it’s very powerful.”

Top photo: Thomas Baucom, Credits: Superhuman

The story of Leaping AI’s first customer is a Silicon Valley classic. The company, which creates voice AI agents for sales and customer service, was part of the famed San Francisco startup accelerator Y Combinator in 2025. Leaping AI presented at the accelerator’s demo day, and following the event, got its first message from a potential customer. 

After the initial interest, co-founder CEO Kevin Wu said it took a lot of meetings to actually close the deal with the interested home-remodeling company. Like any early-stage startup, Leaping AI was unknown, and there were other more established players in the space the customer could’ve gone with.

To convince them, he had to keep selling, narrowing in on a reason the company should choose him over the more established competition. What ultimately convinced them was his pitch that Leaping AI would be a very small and nimble partner, and they would work directly with the company to build custom integrations the bigger players wouldn't do. 

“I turned being small into an advantage by saying, ‘hey, I would actually work with you on a closer level and be a better partner,’” he said. 

The customer signed on, and Wu said they got right to work to deliver and soon reduced the company’s call center costs by around 60%. Furthermore, the company later referred two other customers, showing the power of the flywheel and making good on the promises that earn a startup its first deal. 

For Wu, it was his first startup, and closing that deal was unlike anything that came after. 

“It's a very high job, I think, but it is very satisfying,” he said. “Obviously, you deal with a lot of setbacks, but at some point you get used to it. It's part of the game.”

Some founders change their location on LinkedIn to larger cities to improve their reach, but not Rémi Louf. Not only does this French founder live in a village a 70-minute drive away from Paris, but he rents office space in the local castle. This, too, can attract attention — and convey the image of an entrepreneur who doesn’t follow the flock.

Louf’s office setup is the pragmatic decision of a fully remote CEO: he lives with his wife and kids in one of France’s prettiest towns, whose picturesque castle happens to rent out coworking space. It doesn’t have A/C, but it has a fiber connection, and is “cheaper than a garage in Palo Alto or an office in Paris.” But the bargain has other benefits.

“When I was in SF a couple weeks ago, I got recognized by X users as ‘the guy from the castle’,” he told The Infinite Loop with a laugh. Louf has been humble-bragging about his ‘Frenchmaxxing’ lifestyle on X, and in a July first-person account for Business Insider titled “My startup has a literal moat, thanks to the French castle I run it from.” 

Credits: Screenshot Louf’s X account.

The startup in question is Dottxt, which forces AI models to follow a specific structure in their answers. But the moat may be Louf’s willingness to sail alone. It takes a contrarian mindset to be early to a problem, but also to keep on going after it looks solved.

From upstream to mainstream

The likes of Anthropic and OpenAI now also provide tools to control LLM output, but there were none three years ago when the team behind Dottxt launched open-source library Outlines to make it easier to adopt AI in production. At a time when developers still had to resort to prompt hacking, they used their background in statistics to make AI results more readily usable in workflows — for instance, by turning customer emails into standardized support tickets.

“For a year and a half the field was skeptical. Then everyone copied the surface. Structured outputs became table stakes,” the team wrote. While this made their open-source project very popular, the rising tide could have drowned the startup they had founded in 2023; being first doesn’t always pay off. 

Louf acknowledged it wasn’t smooth sailing, and being challenged on its initial use case — data extraction — rocked its boat. “Last year was quite rough,” he said. That changed with the rise of agentic AI, which opened up a double use case for Dottxt: give agents a structured format to communicate with each other and to take action in the real world through function calling, also known as tool calling.

According to Louf, tool calling is only solved on the surface: it is seamless with closed-source models, but not with open models, hindering their adoption despite growing interest in the alternative they represent. To address this, Dottxt has built a new product called dotlambda, and says it also improves token efficiency, a rising concern as AI moves from training to inference.

Surfing the inference wave

The concomitant rise of agents, open models and AI moving into production have turned the tide for Dottxt. “We are seeing a lot of traction around our tool calling product,” Louf said. The startup has bounced back since February, with strong interest from inference platforms operating in the cloud and on device. “We were a bit early, and the market is catching up, which is nice.”

Dottxt may have been early, but it was also lucky: its approach turned out to be better suited for inference, Louf said. Compared to other methods that affect performance, and therefore costs, he thinks Dottxt is better placed to surf this new wave. But Louf being French, his optimism is tempered. “I don’t think anyone has a moat,” he said. 

In his view, this also applies to Dottxt; a player with infinite amounts of money could copy its technology. But for Louf, the key is time: he estimates that the startup is already “one year ahead,” and says its team is both talented and moving fast. By the time others can catch up, he hopes it will already be further ahead.

Looking at it differently, this combination of talent and velocity may qualify as a moat, a concept that investors have started to redefine more flexibly for the AI era. Dottxt’s funding suggests it met their criteria: it raised $11.9 million to date, via a $3.2 million pre-seed round led by deep tech VC firm Elaia in 2023 and an $8.7 million seed led by EQT Ventures in 2024.

What was once his intuition is also now backed by facts: “Everyone sees the same data as we do; the use of open-source models is really on the rise.” If this says anything about his ability to spot trends earlier than others, don’t be surprised if it becomes a thing to live the castle life — or at the very least, to work wherever you please. 

Top photo: Co-founder and CEO of .txt Rémi Louf. Credits: Louf’s account on X

Nucleo calls itself the first agentic platform for oncology. What its founders mean by the word is an agent that coordinates specialized imaging models within a defined workflow, while clinicians retain the final say.

Founded in 2025 by mathematicians Angelica Iacovelli and Luca Pegolotti, both trained at Politecnico di Milano, the company analyzes CT and MRI scans to extract body-composition metrics and handle tumor detection, segmentation, and classification. It began with lung nodules and is broadening from there. Nucleo says its automated analysis runs 2,500 times faster than manual segmentation.

The pair's route into the problem ran through conversations with clinicians at Stanford, where Iacovelli was modeling cardiovascular blood flow and Pegolotti was researching at the School of Medicine. "We were talking with a lot of doctors to understand their workflows," Iacovelli says, "and we realized that there was a big problem in radiology."

Nucleo’s co-founders talked with The Infinite Loop about how they found their niche and what’s on the horizon.

TIL: You call Nucleo agentic. To what extent are your agents autonomous?

Pegolotti: The reality is that the clinician always has the final word. We provide an indication of what the image looks like, what data we can extract from the image, but in the end it’s the clinician who decides what to do with the data that we provide.

Iacovelli: Being agentic doesn’t mean that we decide everything. We automate a bunch of tasks. At the same time, the final report is reviewed by the doctor, who decides whether they agree with our findings.

TIL: So what is the agent actually doing?

Pegolotti: The AI models that we trained ourselves, or that we make available, are exposed to an agent. We have some workflows that require different AI models to interact with each other. Some of the models are more classic, like computer vision-based. Some are a little bit more cutting-edge, let's say. The latest ones, for example, are based on vision transformers for segmentation. But the agent handles the orchestration. We have different models, and the agent decides which one to call based on the workflow.

TIL: What sets Nucleo apart from existing clinical imaging tools?

Iacovelli: We started with a specific wedge product that enables us to provide a total body composition assessment of the patient, an approach no one has taken so far. In addition to the data points for a specific tumor, it also provides a complete nutritional and metabolic picture of the patient. We’re also building a framework that enables us to generalize across different tumors. We don’t want to build classic computer vision models that can detect a single tumor, a point solution. We’re building models that can generalize across a lot of different diseases.

TIL: You describe a framework for different diseases. How far beyond cancer does it go?

Iacovelli: It’s already expanding. We work with oncologists, radiologists and surgical oncologists, but we also work in obesity and weight loss, where body composition assessment is really important, as well as in geriatrics.

"A really big barrier to entry"

TIL: What has been hardest since you started Nucleo?

Iacovelli: Something that all healthcare companies struggle with is how slow everything is. Startups, by definition, need to be really fast to be compatible with the timelines of VC funds. At the same time, we operate in a space that is really slow. Sales cycles can take years, and it’s a heavily regulated space. You really need to have the resources and the time, and try to be as fast as possible. There’s a really big barrier to entry in this space.

TIL: How did you get from a first conversation with a hospital to a deployment?

Iacovelli: What has enabled us to talk with all these hospitals has been having the first few champions, doctors who have believed in us. The next most effective thing is doctors talking to other doctors. Warm intros or word of mouth work way better than doing cold outreach to a hospital or an institution you’ve never worked with.

“Not something I would make a movie of”

TIL: We run a newsletter format called The First Customer, where we ask founders about the moment a company got its first paying customer. So, plainly: who was yours, and how did the deal happen?

Iacovelli: We can’t disclose the names of the hospitals we work with, but we can tell the story. The first customer we onboarded, the doctor we met, [came] through an introduction from an investor. A lot of doctors we have met have been through investors. This doctor was really excited about the technology. We didn’t have much at that point. But he was super excited, and initially we had this relationship where we were building [the product] following his recommendation. Every startup initially needs to have a design partner, and they were our design partners. They still are. By creating that relationship with a hospital, at that point they will trust you. They will know that you're fast, that once they have a problem, you're always there to answer. And once that trust is built, it's way easier to sell.

TIL: What convinced them?

Pegolotti: I think they saw how we operate. What really makes a difference at this stage is that you have to show that you’re quick at iterating. If they have a specific request, you come back a week later having implemented that feature, and that really makes a difference. We have a background in mathematics. So sometimes it’s a little bit hard for us to understand, does this even make sense from the clinical perspective? Having these interactions with the doctors is very valuable.

Iacovelli: You need to imagine doctors working with these really old tools and interacting with big vendors that are really slow. So when they work with fast-paced companies, they’re like, “wow, this is fantastic.” It's really important to be fast and to show that if they have a problem, we're really quick, and we can build something incredible in a few days. Which is something they're really not used to, working with big players.

TIL: Was the close itself spectacular?

Iacovelli: It was not spectacular. It was not something I would make a movie of.

Pegolotti: It was almost all virtual. We’re often in the US and Switzerland, and it’s difficult sometimes to just meet in person. 

Iacovelli: We always try to go in person for training or to create a relationship, but in this case the contract was closed virtually.

TIL: You won first place in the Medical Imaging category at the Nebius AI Discovery Awards. What’s your plan with the credits?

Pegolotti: We believe the future of radiology will not be built from hundreds of disconnected AI models, but from unified foundation models capable of reasoning across imaging, clinical history, and multimodal patient data. We will use the Nebius compute to lay the groundwork for these models, starting with oncology foundation models for pan-cancer detection, lesion quantification, and classification.

Disclosure

Iacovelli and Pegolotti spoke with The Infinite Loop at the Nebius AI Discovery Awards ceremony in London, where Nucleo took first place in the Medical Imaging category. The Infinite Loop is a Nebius editorial project. Editorial decisions are made independently.

Philip Shoemaker, founder and CEO of new identity rights and deepfake detection platform PersonaShield, is as excited about his first customer as any entrepreneur can be. That’s not just because the deal validated his new venture, but also because the customer is someone he, “an old skater,” has personally been a huge fan of for decades: global skateboarding sensation Bob Burnquist. The craziest part about it? The deal started with a “like” on LinkedIn.

One day while scrolling, Shoemaker happened upon a post from a friend of his, and he saw Burnquist had “liked” it. He reached out to the friend asking, is that really THE Bob Burnquist? It was, and it turned out that the friend and Burnquist were real-life close friends. Luckily for Shoemaker, the friend was willing to make an introduction.

“Bob wanted to meet me because of my background at Apple, and I started advising one of his companies, and he started advising us on features for [PersonaShield]. So it was kind of a you-scratch-my back-and-I'll-scratch-yours-type of scenario where we're both helping each other out, because we're entrepreneurs that are trying to make things happen,” Shoemaker said. 

Philip Shoemaker, founder and CEO of PersonaShield

At that time, PersonaShield was still trying to figure out if it was purely a takedown company or would focus on image generation with safeguards.

During a conversation with Shoemaker, Burnquist expressed he was feeling really worried about deepfakes, specifically people using his face to falsely endorse products to manipulate his fans, especially his many Brazilian fans who are economically disadvantaged. But at the same time, he was stepping into a role heading up the new X Games League, and he was thinking it would be great if people could generate images of the competitors as a new form of fan engagement.

Hearing Burnquist’s nuanced thoughts about how he wants his likeness to be used made clear that unauthorized deepfakes and authorized usages are all part of the same problem — and would need to be part of the same solution. 

Read the full story on how PersonaShield is trying to build her product here.

Top photo: Image generation sample for skateboarder Bob Burnquist. Credits: PersonaShield

Can you trust a celebrity’s endorsement? The question is no longer just about whether you trust their judgment; it’s about not knowing if they ever really made the endorsement in the first place. 

As AI models generate increasingly hyper-realistic images and videos, it’s causing an explosion of deepfakes and impersonation scams involving public figures. Global deepfake scammers made off with $96 million in just the first three months of 2026, with the majority of successful scams hinging on imitating government officials or celebrities to endorse fraudulent investment opportunities, according to SurfShark. 

Now, a new startup, PersonaShield, is launching to not only help public figures stamp out deepfake scams, but take greater control of their likeness overall in this new AI world. The company is building infrastructure around “AI identity rights,” helping creators, influencers, celebrities, executives and public figures manage how their image and likeness are used across AI systems and social platforms. 

“Every creator we've spoken to has said, ‘well, I want people to be able to generate images of me. I just have certain requirements,’” said founder and CEO Phillip Shoemaker.  

PersonaShield’s service allows public figures to define what uses of their likeness are permitted, automate enforcement around misuse, and even create pathways to monetize their likeness through authorized uses like fan art and online content. For example, an executive could ban all uses of their likeness, while an influencer could set terms to allow their likeness to be used in certain ways for a licensing fee.

“We give guidance, but everything's up to them to choose how they protect it, if they want to protect it, and if they want to monetize it,” said Seamus Hennessy, product manager at PersonaShield. 

There are already various startups focused on detecting deepfakes, a crowded field of take-down services, and some creative solutions for designating authorized content. With PersonaShield, Shoemaker is looking to combine both sides of public figures’ modern identity concerns, bringing together protection and marketplace. 

“We're about owning your likeness, and protection is one aspect of it,” he said. “But image generation is another aspect — image generation with monetization and paying the people whose images you're generating from.” 

Choose your own (likeness) adventure

As Apple’s former head of app review who helped create app store guidelines, and later an executive and advisor in the decentralized identity space, Shoemaker knows a thing or two about managing scams and fraud online. He also knows about how quickly enthusiasm to monetize on a new technology opportunity can rise.

In addition to servicing public figures, PersonaShield also serves as a marketplace. People can go to the PersonaShield website, browse and search for celebrities or influencers, and purchase permission to use a public figure’s likeness however they want — so long as the use falls within the person’s set guidelines. This isn’t for using their likeness in, say, a feature film, but rather for fan engagement, online content, memes, and the like.

Image generation sample for skateboarder Bob Burnquist. Credits: PersonaShield

PersonaShield’s first customer was professional skateboarder Bob Burnquist. Shoemaker described Burnquist as making it very clear that if he’s in a skate competition, he needs to have Banco do Brasil on his shirt. Because it affects his sponsorship, it’s his absolute requirement, and he can't have people generating images of him without it.

Combating AI with AI

For PersonaShield, AI is the problem and also a large part of the solution. Like most startups, it’s using AI widely throughout the company for everything from engineering to marketing. What’s more, AI is also a core part of the company’s service and technology infrastructure underneath its image analysis capabilities. 

“We use AI to tell us, is this an image of Snoop Dogg? And if so, what's he smoking? Oh, it's not cannabis; it's crack. Okay, that falls outside of his safeguards,” said Shoemaker as an example (Snoop Dogg is not currently a customer).

Hennessy compared the current methods for deepfake detection used throughout the industry to a game of whack-a-mole, largely dependent on analyzing pixels to classify images as deepfakes in a more strict, deterministic way. With PersonaShield, he said they’re trying to bring more context and understanding about the specific person into the detection pipeline. This includes analyzing what they’re known for, what they’re often associated with, and what’s counter to that, and overall, building a profile of the person and using that to flag content that falls outside their set safeguards. 

“We're trying to think about, how can we make this more adaptable? So no matter how accurate these images get, it can almost think for itself in terms of like, I know who this creator is, what content they promote, what they tend to stand for, what they say, who their sponsors are, where they’re known to be, and what they tend to promote or not,” Hennessy said. 

PersonaShield’s ambition for the next few months is to use AI to automate takedown requests for unauthorized content. This is currently an incredibly tedious task, which involves manually reaching out to each platform and requesting content to be removed (and following up and escalating if it’s not). The idea is for agents to handle it on the backend without any manual interaction. 

PersonaShield founder and CEO Phillip Shoemaker. Credits: PersonaShield

Speaking about the rising deepfake issue, Variety reports that Hollywood agents (the human kind, who represent celebrities) say the previous manual methods for pursuing takedowns are insufficient for the sheer number of violating posts they’re seeing because of AI. Agency sources describe meeting with “dozens” of startup solutions and told Variety they expect — and are already seeing — more talent lean into these kinds of automated, all-in-one solutions. Other choices include Loti and Vermillio, which also help facilitate licensing opportunities and scans the internet for unauthorized content to automatically issue takedown requests.

Overall, PersonaShield wants to put people in the driver’s seat in this new AI era. To the team, this means empowering people who feel like the problems associated with AI and deepfakes were thrust upon them without any say. As Shoemaker sees it, the frustration and anxiety come from not having a choice in any of this.

“So many people are angry with [AI] because it's all trained on our data without even asking us,” he said. “We want to move forward and start asking the questions, like, do you want to allow images to be generated of you? What do you want generated, and what don't you want generated?”

Top photo: Takedown request sample for skateboarder Bob Burnquist. Credits: PersonaShield

If you have walked into an Aldi or a Tesco recently and picked up a bottle of water without going through a checkout, cameras in the ceiling were quietly reconciling what your bill should be. It's a world Thomas Knox knows well. He helped put ambient sensing into more than a hundred autonomous stores, rising to head of product at the San Francisco computer-vision company AiFi.

VitVio, which Knox founded and now runs as CEO out of London, does something similar to what AiFi does. Cameras and sensors in the ceiling. Models tracking who and what are in the room. Only, this time, the equipment goes into the operating room instead of supermarkets.

VitVio has a publicly announced clinical collaboration with the Royal Orthopaedic Hospital in Birmingham, and Thomas Knox says the company has expanded to hospitals across the UK, US and Europe. Thomas Knox and VitVio’s chief scientist, Grzegorz Jacenków, sat down with The Infinite Loop to talk about how their system aims to simplify the lives of staff and how the tech is far from the hardest part when working with hospitals as your customers.

TIL: What’s actually happening when your product is in action?

Knox: At VitVio, we build ambient sensing and computer-vision tech. We go into the operating room, deploy cameras and sensors, and track who’s doing what, when, where, and why. Then we use that to take over administrative work and drive efficiency across the OR block. What we’re really doing is understanding the context: the stage of a surgery, patient arrival, induction, knife to skin, wound closure, writing those events straight to the electronic health record. And because we already have a deep 3D understanding of the room, we can recreate the surgery in 4D and replay it. Put on VR goggles, stand next to the table like you were there. The goal is for VitVio to feel like another staff member in the room. One without a physical presence, that you can talk to in natural language.

Jacenków: Instead of running models that just look at the scene, we recreate what’s happening in 3D. We know where every person is, where every tool is, and we plug new [specialized] models onto that representation. If a patient has arrived, we can notify the next team to prepare the following one. If a surgery is close to finishing, we bring cleaning staff in early. We are a native 3D company.

“If you add another step to my process, you’re going to end up on the table yourself”

TIL: Hospital ROI and reduced burden on the OR team can be in tension. VitVio is pitching both. So, what’s the actual goal?

Knox: Both, and they have to be equally balanced. Everything we build has two objectives. One, make the staff’s life easier. Two, deliver hard-dollar ROI to the hospital. If you make the hospital more money but the staff’s life gets harder, you are harming the hospital in the long run. Staff burn out. Throughput drops. When we were first starting the company, we were talking to a surgeon, and he said something that became a motto: “If you add another step to my process, you’re going to end up on the table yourself.” Every part of our system has to eliminate an existing step or translate it into something easier. If we are adding a step, it’s a no-go.

TIL: What has been the hardest thing to overcome?

Knox: The tech is not the hard part. I have been part of three unicorns. Our team has come through big companies. Greg was at Amazon. I ran product at AiFi. If you walk into the O2 Arena in London, that is our tech. It runs at Aldi, Tesco, Dollar General, Marks & Spencer. The tech, we understand. The biggest challenge, and it is not just us, is that hospitals are cautious animals. Rightfully so. They are dealing with our health. But what we are proposing is not a step function; it is a leap in how they operate. It changes the entire way a hospital runs. Only the best-of-the-best fully understand where it is going. Everyone else follows, but cautiously. You have to bring them somewhere it is already working before they believe the future is real.

Jacenków: We are also privacy-first. When cameras appear in an OR, people can be afraid of what’s happening. The moment a feed leaves the OR, everything is anonymized. We do not care who is doing what on a per-person basis. And our servers are installed on-site. Data stays in the hospital. We send them notifications about the next step in the process. That is it.

VitVio’s chief scientist, Grzegorz Jacenków. Credits: VitVio

Knox: Previous versions of this technology focused on individual metrics, and that gave a Big Brother feeling. We do not track anyone’s bathroom breaks. We track the process. But we have to say that upfront. Because if I don’t tell you anything except that there are cameras and sensors in the room, your brain runs off with ideas about how it could be used against you.

TIL: Walk us through the first deployment. How was it?

Knox: Simpler than people imagine. We come in for a site visit, take out an iPhone, and do a LiDAR scan of the operating room. Less than two or three minutes. The system spits out a camera plan with the mathematical overlay of coverage. We hand it to the hospital’s team to install [the equipment] with the exact measurements. The whole install takes about four hours, overnight. The cameras aim and calibrate themselves. Everything else runs off-site. Our first deployment was up in four or five hours. It hasn’t been touched since. Compare that to the autonomous stores, where we had to redeploy 30 times because things break in the chain and it is not obvious where. Here, we had the experience already. 

And we're moving into the territory now where the hospitals will deploy it completely themselves. They don't actually need us involved. So, the very first one, we'll come in, we'll white-glove it, we'll work with the team, we'll show them exactly how to do this right. But after that, we literally hand them a mobile application for an iPhone so they can do the LiDAR scan themselves, and everything else is just done through the app. All they have to do when the server arrives is just plug it in.

TIL: Is hospital self-deployment just a vision, or is it near-term?

Knox: It is already in the process. When you deploy hardware, you are almost linearly scalable. Your deployment team becomes the biggest team in the company. We wanted to grow without a two-thousand-person deployment team. So we automated it for ourselves first. Now we're packaging our systems into a tool that we're handing to the hospitals within the next 6 to 12 months or so.

TIL: You've picked up second place in Digital Health at the Nebius AI Discovery Awards. What’s your plan with the credits? 

Jacenków: We are incredibly grateful to have been selected for second place in the Digital Health category, particularly from a global field of nearly 650 highly competitive submissions. We already develop specialized models designed to understand the operating room and its workflows. The Nebius credits will allow us to step up the scale, pace and ambition of that research. Our ambition is not simply to adopt advances made elsewhere, but increasingly to contribute advances of our own.

Disclosure:

This interview took place on July 1, 2026 at the Nebius AI Discovery Awards ceremony in London. The Infinite Loop is Nebius' editorial project. Editorial decisions are made independently.

Top photo: VitVio’s CEO Thomas Knox. Credits: VitVio.

When Bobby Healy founded Manna Air Delivery  in 2018, he had little idea of who his first customer would end up being, or the extreme conditions that would lead to the partnership.

Covid-19, which sent the world into lockdown in 2020, created a new and urgent use case for drone delivery: getting medicines to people in remote areas with minimal human contact.

“Our first customer partner became the Health Service Executive (HSE), Ireland’s national health service, and we were delivering prescription medicine there in rural areas,” Healy said. “[We had] technology that can deliver things hands-free, and they were looking for innovation to say, ‘How do we move critical goods around?’”

For Manna, which intended to target the food delivery market, delivering medicines created a new layer of regulatory challenges on top of the airspace safety requirements it was already working on.

“We’re regulated by aviation regulators, but there’s a whole other world of clinical regulation that comes in around the ‘chain of control’, keeping medicines at the right temperature,” Healy said. “We decided to partner with other companies to do the clinical chain of control side, and all we need to provide is the box that they put their box into.”

Working through these challenges wasn’t easy, but Healy said that the experience showed the impact that drone delivery can make.

“Elderly patients that couldn’t leave their houses swore by it,” he told The Infinite Loop. “They were able to get a video consultation with their doctor, the doctor was able to issue a prescription electronically to the local pharmacy. We were able to fly it to the house, so there was zero human contact and zero need to leave the house.”

Beyond delivering potentially life-saving medicines, Healy said that working with the HSE served as a proof of concept for the company in another way.

“We were able to get people there in their 80s to adopt a technology like this, which speaks volumes,” he said. “It's not just young people and techies adopting this technology, because it's so easy to use and it's so meaningful for people that are incapacitated or have mobility issues.”

For more on the tech and social issues facing companies like Manna Air Delivery, read the full article here on The Infinite Loop.

ALSO THIS WEEK ON THE INFINITE LOOP

What happens when a €5.5B French insurer builds an AI companion its members trust with their symptoms, and puts Kylian Mbappé on the team to nudge them out for a walk. Anna Heim on Alan's bid to move health insurance from claims processing to prevention. Read the piece.

In the summer of 2018, Bobby Healy was sitting in his garden in a Dublin suburb enjoying a glass of wine, when he decided he wanted nothing more than a bag of fries to go along with his drink. But there was a problem.

“It was impossible to get a delivery here on the weekend, or if the weather’s bad,” he said. “There’s just not enough drivers to work for delivery companies, so you could be waiting an hour for your bag of chips.”

Rather than just settling for a piece of toast, Healy did what any self-respecting serial entrepreneur would do, and set about founding Manna Air Delivery, a drone delivery company that has completed more than 250,000 jobs to date.

As of July 2026, Manna is delivering everything from takeout food to medicine and defibrillators, as it takes on the $180 billion last-mile logistics market, aiming to make deliveries faster, greener and more affordable.

But turning flying delivery robots into reality means fitting powerful AI systems onto small aircraft while meeting stringent safety and regulatory requirements. There are also societal barriers to overcome, as some fear a future of skies crowded with noisy, privacy-invading drones, raising questions of whether more technology in the world is always a good thing.

‘We have to optimize the hell out of the stack’

Healy described how the drone delivery sector has walked a long road to proving that the idea is viable, both to regulators and the companies that use these services, and that the US market has been the fastest to adopt the technology.

“The top companies: ourselves, Google Wing, Amazon Prime Air and Zipline, we're all now regulated and ready to grow,” he said. “We’re live in about 40 locations this year in Oklahoma and Texas, and our friends at Google and Amazon are the same, so collectively there'll be about 10 million deliveries this year, nearly all in the USA.”

Manna’s drones are supervised by human operators on the ground, but use autonomous obstacle detection to fly safely. Healy said the system relies on streams from multiple sensors, processed by a mix of computer vision and more deterministic AI systems, all running on onboard compute.

An engineer works on a Manna drone. Credit: Manna Air Delivery

“There would never be a scenario where you would have anything safety related needing to go off the aircraft and rely on a network, because networks can never be reliable,” he said. “The compute has to be on the edge, which is a difficult challenge because of the energy it consumes. We have to be sacred with our battery and energy, and also weight on the aircraft.”

Healy said that Manna has achieved this by relentlessly fine-tuning both the software and hardware on its drones to deliver maximum compute with minimal weight and energy use: “We have to optimize the hell out of the stack.”

When a pizza becomes a wing

Other drone delivery companies are solving more item-specific challenges. Flytrex, founded in Tel Aviv in 2013, focuses solely on food delivery. One of its early partnerships with Little Caesars in Dallas created an engineering problem: the drag of a large flying pizza.

“Delivering large pizzas has been a big challenge for us, because now you're flying with this huge sail,” said Amit Regev, co-founder and CEO at Flytrex.

The company has tackled this by developing a system that holds the pizza at a consistent angle to minimize drag while in the air.

“It’s designed to keep the pizzas at about five degrees. When you’re flying forward, you’re actually getting almost a wing shape, so rather than having a sail, you’re almost generating lift,” said Regev. “With that design, when you do endurance testing, you can actually fly even further with the pizza than without it.”

A Flytrex drone carrying a 16-inch pizza. Credit: Flytrex

Flytrex is also delivering sandwiches for New Jersey-founded chain Jersey Mike’s and signed a partnership with Uber Eats in September 2025. Regev said the sector has now moved from proof of concept and regulatory approval to making the unit economics work at scale.

“How do you get drones, which are sophisticated devices with a lot of technology, to deliver burritos at $5 and actually have enough margin to be profitable and run a healthy business? My personal opinion, just by looking at the competition, is that nobody’s solved the economics yet. I think we’re getting very close,” he said.

Regev said a big part of bringing costs per delivery down involves using AI to improve the management of Flytrex’s fleets.

“Optimizing drone routes is not as easy,” he said. “We have a huge team that deals just with how to get the most efficient route.”

Concerned communities

Beyond optimizing fleets and onboard compute capacity, drone delivery companies have another challenge on their hands: winning over communities who are suspicious of the technology. In June 2026, Texas residents protested outside Amazon’s drone facility, to voice concerns over noise, safety and privacy, following demonstrations in Georgia and Australia.

Healy believes some people’s concerns over noise pollution are based on a negative perception around smaller drones, which make a higher-frequency sound than Manna’s vehicles.

“The typical drones that most people are used to naturally generate really angry, buzzy bee frequencies that we don’t. We're much bigger aircraft, so much lower frequency noise,” he told The Infinite Loop.

Addressing privacy concerns, Healy said that onboard computing means that no data from Manna drones’ sensors ever reaches a cloud server, adding that safety conversations around delivery need to take the status quo into account.

“To make money, delivery drivers have to go fast. Delivery drivers are six times more likely to be involved in a collision than a normal driver,” Healy claimed.

Healy and Regev both believe that drone delivery will lead to a more convenient, greener and safer society, with fewer vehicles on the roads. The key question will be whether companies like Manna and Flytrex can make the unit economics of cutting-edge technology work for the low-margin business of last-mile delivery.

Top photo: A Manna Air Delivery drone in action. Credit: Manna Air Delivery

The 2026 World Cup may give the impression that footballers could advertise anything and everything, but not French soccer star Kylian Mbappé, who has been more selective. This adds weight to his involvement as investor and ambassador at Alan, a French unicorn that provides one million-plus members with health insurance on behalf of their employers.

Now valued at €5.5 billion and operating in four countries, this European health insurer has certainly received much larger checks than Mbappé’s — the latest is a €400 million investment from global tech investor Prosus alone. But France’s captain also became a “teammate” of Alan — and of AI — in its goal to put prevention front and center.

With an in-app challenge called “Move with Mbappé,” and with his endorsement, Alan encouraged its users to follow the athlete’s footsteps and become more active. That game ended when Mbappé met the winner before taking off for the U.S., but it reflects Alan’s use of technology to align incentives and nudge its members in the right direction.

Screenshot: Alan

Gamified nudging is also how Alan landed Mbappé as an investor. “His mother, Fayza, was using Alan, [walking] more steps thanks to us, so they got to know us,” its co-founder and CEO Jean-Charles Samuelian-Werve said in a talk attended by The Infinite Loop during the South Summit Madrid conference in early June.

Amid mounting costs, shifting from treatment to prevention has become a necessity; but Alan still hopes to move faster than the incumbents it joined when it received its health insurance license in 2016 — a first in France since 1986.

Running fast

“The idea behind Alan is that health cannot wait — health cannot wait for you to find an appointment to the doctor, you to get reimbursed, you to act on the small problems that could become very big if you don't cure it,” Samuelian-Werve said on stage.

Many health insurers have now realized that AI can be part of the answer, but Alan was arguably early to the game. Samuelian-Werve and his co-founder Charles Gorintin, Alan’s CTO, are also non-operating co-founders and board members at French AI champion Mistral, and didn’t miss that machine learning could help process reimbursements and queries faster.

For Alan, automation unsurprisingly translates into significant savings — €300,000 for just one month, Samuelian-Werve disclosed on LinkedIn. This also comes at a cost — many of the lower ratings of Alan on Trustpilot are from users complaining that chatbots had them running in circles when a human agent would have been better suited to solve their case.

Still, the company is doing better than average on the customer satisfaction front, with a reported net promoter score above 68.

A different ball game

Alan wouldn’t be as liked if it weren’t for automation, which helps it answer member requests in 20 seconds on average. That’s also why it has built Mo, which aims to go beyond customer service and insurance-related questions. Now available to all Alan members in France, it is positioned by the company as an AI healthcare companion that can switch topics fast.

Early data gathered by Alan indicates that members do ask medical questions to Mo — often relating to symptoms, or to decision making. That’s also where mistakes could cost lives, which is why Alan has added guardrails. “Every message is verified by a doctor within 15 minutes,” the company stated when it started deploying Mo in 2024. 

The fact that AI alone may not be ready for triage hasn’t stopped doctors and patients from asking health-related questions to the likes of ChatGPT. In January 2026, its parent company OpenAI announced ChatGPT Health. But for Alan's CEO, Mo has an edge “because it knows your history, past claims, family, and past interactions with [Alan’s] medical team.”

Using this context, Mo can also act proactively, suggesting checkups, booking appointments, and building gamified routines with Alan Play based on findings from Alan’s new blood test, Alan Precision. This approach is what keeps members coming back:: in 2025, its app claimed 26% weekly active users and 10% daily active users.

What makes these usage levels manageable at scale is automation. The company claims to have adopted AI as its operating model; for example with Hopper, an internal tool that lets all of its team contribute to the code. With only 850 full-time employees, or “Alaners,” the company now targets €1 billion in annual recurring revenue for 2026.

Players and positions

In a market dominated by non-profits, Alan’s hybrid positioning has raised questions. The countries where it operates — France, Belgium, Spain, and Canada — have health systems where mandatory insurance isn't priced based on health status. But there is still something slightly unnerving about asking very personal health questions to a privately owned chatbot paid for by your employer, even though strict walls are in place to maintain confidentiality.

Since the main privacy-related risk in this context would most likely come from data leaks, some of Alan's reassurance on these issues will stem from which models it uses, where it stores its data, and how it plans to keep it safe. Despite its tech savviness, it may not have much of a leg up, as France’s tsunami of data breaches in recent months also affected one of its providers. 

Its competition is exploring AI too. There’s market leader Vyv, for instance, but also Doctolib. This French unicorn already owns a staggering share of the appointment booking layer, and the overlap with Alan could increase. In a LinkedIn post, Samuelian-Werve summed up a test in which Alan demonstrated that voice AI could help reduce the cost burden of canceled appointments for healthcare providers.

Screenshot from LinkedIn

While foreign LLMs will also go after this market, Alan is preemptively reinforcing its moat for the AI era by embedding itself deeper into the lives of its members, while seeking to align incentives from all sides. “The future of health is built at the intersection of insurance, prevention, care and AI,” Samuelian-Werve wrote. It is hard to see how competitors would disagree, perhaps leaving Alan with no other option than to outrace them.

Top photo: Jean-Charles Samuelian-Werve (on the left), Alan’s CTO and co-founder Charles Gorintin. Credit: Alan

The startup mpathic is building red-teaming AI models and chatbots for safety risks. Today it has customers across verticals, from frontier AI labs to customer service providers and healthcare companies. The company’s very first customer, however, was in the life sciences space, and the deal was scored by CEO and founder Grin Lord herself, thanks to some clever maneuvering. 

The first customer was familiar with Lord’s previous work in fidelity monitoring (measuring how closely a therapist or provider follows the intended treatment protocol) and using natural language processing to evaluate the quality of medical providers and therapists from recordings. They wanted her to consult on their new product. She told them she’d do it — and would even do it for free — but they’d have to go through her new company, and the contract had to include a low-level subscription to their software.

To close that first sale, Lord had to do a lot of education to help the buyer understand the state of AI and what they were trying to do. They also had to get on the same page about how to ensure mpathic could meet the company’s compliance needs. 

This almost tanked the deal: as a small startup, mpathic didn’t yet have the team for the kind of complex compliance this company required. Lord believes they were considering alternatives, including her colleagues from past companies. But the customer ultimately decided they could solve this hurdle through collaboration, and they worked closely with mpathic to ensure compliance and even funded trainings for the team. 

Since Lord previously launched two startups in the natural language space, this wasn’t her first-ever “first customer” moment. But it still had those same high-stakes feelings. First came the validation: “Okay, they get it!” she said. Then came the realization: “Wow, they are going to pay for this!” 

“As the deal went on, it became clear that the scale of what the company wanted to do required both our AI platform and our services to build custom classification and speech models with ground-truth validation. So what started as just wanting to talk to me grew into a small SOW and software contract, and then grew into a five-year, multimillion-dollar relationship,” she said.

From that first customer, mpathic has grown into one of two startups now shaping the field of chatbot safety. The other, Circuit Breaker Labs, runs its red-teaming on autonomous AI agents instead of human clinicians. Both think the other approach hits a wall. Read the piece by Sage Lazzaro on The Infinite Loop.

Photo: Grin Lord, founder and CEO of mpathic. Courtesy of mpathic.

People are increasingly treating AI chatbots as friends, confidants and therapists, and they’re turning to them in the mundane and tough moments of their lives. This phenomenon hasn’t come without consequence though: intense conversations with chatbots have regularly been linked to mental health spirals, and in some cases to suicide and murder. 

According to data shared by OpenAI in October 2025, over a million people explicitly express suicidal intent to ChatGPT every week, and at least half as many show “possible signs of mental health emergencies related to psychosis or mania.” And while the way people are using AI chatbots has been shifting dramatically, according to the 2026 annual analysis in Harvard Business Review, there is one consistency: therapy/companionship has remained the top use case for AI chatbots overall.

A growing field of startups is building safety infrastructure for high-risk chatbot conversations — both for general-purpose chatbots and those designed specifically to act as therapists, AI friends and everything in between. Some, like mpathic, are founded by clinicians themselves, while others, like Circuit Breaker Labs, are being built by rising tech entrepreneurs looking to tackle the problem of their day. Both startups agree that explicit indicators of mental health crises aren’t the issue; a filter catches those. Rather, they’re focused on subtle linguistics and how interactions build over time. 

“What is more of the problem isn't just the one word,” said Grin Lord, a clinical psychologist and CEO and founder of mpathic. “It's the behavioral pattern of what we call a multi-turn interaction, where someone may leave bread crumbs.”

While mpathic is going all-in on human creativity to pressure test models for mental health safety, Circuit Breaker Labs is fully betting on autonomous AI. Both think the opposing approach eventually hits a wall.

The AI approach

Launched in 2025 by the brother-and-sister team Shirali Nigam and Arul Nigam, Circuit Breaker Labs built an agentic red-teaming system that pressure-tests chatbots, hunts for hidden mental-health vulnerabilities, and acts as a check on AI labs’ work.

“For some reason, in this industry, safety is just taken at face value from the developers. And I think a lot of these big companies are taking efforts to make sure that their products are safe, but it's really hard to see your own blind spots,” said CEO Shirali Nigam.

Shirali Nigam and Arul Nigam, Circuit Breaker Labs. Credit: Circuit Breaker Labs

The system is designed for continuous monitoring, running daily on models that are already in use to ensure they don’t “drift,” referring to the tendency of large language models (especially “wrappers” built on top of foundation models) to lose predictability and slip their guardrails over time. The company worked with clinical advisors to understand how mental health conversations emerge in the real world and determine what areas AI should be evaluated on and what remediations are necessary when they fail, said Nigam. They specifically focus on linguistic nuances in users’ language that can slip past AI models’ safety detection systems, testing with slang and differing ways people might express the same sentiment. Even typos pose a huge risk and can easily allow alarming conversations to bypass models’ safety guardrails, said Nigam. 

When the system detects a vulnerability, it automatically goes deeper and compiles insights on the detected gaps. The company then provides customers with regular reports detailing what to fix, usually on a monthly basis. The key, as Nigam sees it, is that the evaluation runs continuously, made possible by the bandwidth of using AI agents.

“Because it's an API that's autonomous, you can run like hundreds of thousands of test cases with it,” she said. “So just in terms of capacity of how much manual red teaming a group could do, versus this, you can look at a lot more breadth.”

The human approach

At mpathic, Lord points to the same issues around slang and dialect, which she says present significant challenges for younger users. mpathic’s red-teamers test models in different English subcultures, in other languages, and across generational and topical registers, she said. In particular, they look for how AI models respond to subtle signals that may not clearly indicate cause for alarm when taken in isolation, but paint a worrying picture altogether. 

“If we can demonstrate a trajectory, a behavioral pattern where the model is missing things over and over again in a short, but not single talk-turn [conversation], that's a concern,” said Lord, with “talk-turn” referring to each time either the human user or chatbot responds. 

But rather than AI, the company — which was founded in 2021 — is relying on over 5,000 curated experts to do so, most of whom are licensed clinicians. Lord described her human red-teamers as “endlessly creative” and working almost like actors, constantly coming up with scenarios you wouldn't think of but that are entirely plausible. Some work in a more methodical way, developing a plan to build a rapport with a model and then push it in specific ways. However, she said the most revealing discoveries are like “pulling on the thread of a sweater” — when you’ve already gotten to know a model and then it starts behaving in a totally unexpected way. Ironically, the most surprising behaviors they discover come up when people try to improve a model, and then you find out that a behavior that was solved for is now a problem again. 

CEO of mpathic Dr. Grin Lord. Credit: mpathic

mpathic is starting to use AI for red-teaming in minor ways, which Lord described as “kind of cloning our creative folks on things that are more known, and that we've repeated over and over.” But she thinks this approach has serious limits, and that humans who can adaptively red-team are required. 

“I think the problem with that…is it just runs out,” she said, referring to agentic red-teaming. “You can only do so much with an auto eval or an eval set. You've done enough training, and then you kind of need the next piece of data. They're not infinite.”

The disagreement between the two approaches has no clear answer — and maybe there isn't one. Maybe this is an area where AI and humans need to work together.