Healthcare has become one of the most persuasive arenas for artificial intelligence, but also one of the least forgiving. In this third article drawn from companies presenting at NVIDIA GTC in San Jose, the focus shifts to start-ups navigating health benefits, drug discovery, scientific reasoning, and clinical visualisation, where the consequences of uncertainty are immediate and difficult to correct.
Artificial intelligence in healthcare is often framed as a problem of capability, more data, better models, faster insights. That framing misses where the real friction lies. The difficulty is not simply analysing information but making it usable at the moment decisions are made. Healthcare systems are dense with data, yet patients, clinicians, and researchers still operate with gaps in visibility, clarity, and confidence.
That tension runs through this group of companies. None is attempting to impose intelligence onto a clean system. Each is working where the system itself fails, where patients cannot understand what they are paying for, where drugs fail late despite promising science, where experimental knowledge disappears, and where clinicians still lack clear visual guidance inside the body. The challenge is not simply generating answers. It is recovering context.
Making benefits intelligible
Healthee focuses on one of the opaquest layers of the healthcare system, how people access and pay for care. Its platform centres on Zoe, an AI assistant designed to explain benefits, estimate costs, guide treatment decisions, and help users navigate complex insurance structures in real time. The company also supports open enrolment, provider selection, and care navigation through a single conversational interface. Its website positions Zoe as a continuous layer across the employee experience, bringing together plan data, provider networks, employer programmes, and claims logic into a single view that can be accessed on demand.
The starting point is not technical, but structural. “Health benefits are supposed to be simple, easy to navigate, and affordable, but in reality, they are none of those things,” Ron Zionpour, Chief Technology Officer at Healthee, says. “Most people have insurance, but they do not know how to use it or what they are going to pay. If you go for an MRI, you might pay a few hundred dollars or well over a thousand, even though it is the same machine and the same scan. There is no luxury version of an MRI. The problem is not the procedure; it is the lack of transparency.”
That lack of transparency creates a layer of friction that sits between patients and care. “Our goal is to use AI to make access to a healthier life effortless by putting the right knowledge in people’s hands,” Zionpour says. “Zoe is available all the time, and you can ask questions in natural language. She understands your health plan, your network providers, your employer programmes, and she can show you what your options are, what you will pay, and how to move forward. You can ask about treatment, compare providers, or schedule care, and the system will guide you through it step by step.”
The technical challenge sits beneath that interaction. “The way people use products like this has changed,” Zionpour explains. “The input is conversational, unstructured, and fragmented, so traditional analytics do not work. We needed to extract a semantic layer from that data, and that is where NVIDIA NeMo and NeMo-Tron became important. We use them to build a source of truth across all interactions, and that drives personalisation, behavioural analytics, and how Zoe improves over time.” The result is a system that does not just respond to questions but learns how people navigate uncertainty.
Intervening before the clinic fails
TranscriptaBio is working much further upstream, but it is addressing another form of systemic failure. Its focus is clinical translation, the point at which promising therapies fail to deliver once they reach patients. The company has built a platform around the transcriptome, combining a disease signature atlas derived from single-cell RNA sequencing with a proprietary drug-gene atlas mapping how compounds affect gene expression. On top of this, it has developed Conductor AI, a set of machine learning models designed to predict how new molecules will influence disease-relevant pathways.
Christopher Moxham, Chief Executive Officer and Co-Founder of TranscriptaBio, frames the issue directly. “The crisis in pharma is not about designing molecules,” he says. “We can design molecules that work in theory. The problem is what happens in the clinic when those programmes fail. That failure costs time and capital, but more importantly it delays effective therapies for patients. AI has been applied across drug discovery, but it has not fundamentally changed that outcome.”
The company’s thesis is that gene expression provides a more direct route into the problem. “The transcriptome is the most information-rich way to understand both disease and drug action,” Moxham says. “We have built a disease signature atlas using single-cell data from real patients, and a drug-gene atlas with over a billion experiments that shows how compounds affect gene expression across the druggable target space. That allows us to ask a different question. Instead of starting with a molecule and hoping it works, we start with the mechanism of disease and identify molecules that can address it.”
That shift changes how discovery is structured. “We front-load the clinical translation question,” Moxham explains. “We identify disease mechanisms, match them with compounds, and we know this works because we have already taken molecules into patients and seen them reverse disease phenotypes with clinical benefit. The models built on top of this data extend the approach. Conductor AI allows us to predict novel compounds that tune gene expression in the right way, then we deconstruct those hits to understand which structural elements drive activity. From there we can screen for molecules with similar features and generate new candidates.”
The platform is designed to handle both simple and complex cases. “We can target a single gene with strong genetic validation, or we can work with complex diseases where the signal only becomes clear at the level of gene expression,” Moxham says. “In both cases, we are identifying molecules and targets in a way that is far more capital efficient than traditional approaches.” The ambition is not incremental improvement, but a change in how therapeutic risk is managed.
Capturing what science forgets
Labtree addresses a different gap, the loss of experimental knowledge within scientific workflows. Its platform is designed to capture and structure the context around experiments, turning fragmented lab activity into a usable knowledge base that can support reasoning and decision-making over time. The company describes this as infrastructure for scientific reasoning, built on capturing metadata, protocols, notes, and outcomes directly within workflows, then structuring them into connected knowledge systems that improve as more data is collected.
“Most experimental knowledge is never properly recorded or published,” Hanna Luniak, Chief Executive Officer and Co-Founder of Labtree says. “That means critical context is lost, and in biology this contributes to a reproducibility crisis where a large share of results cannot be replicated. That delays new therapies and increases the cost of research.”
The issue is not simply missing data, but missing reasoning. “Agentic AI has increased the speed of computational research, but it does not capture the intent behind experiments,” Luniak explains. “The challenge is translating human scientific reasoning into something machines can use. That is why we focus on capturing context at the moment decisions are made, not after the fact.”
The system is structured to follow that process. “We capture experimental nodes, metadata, protocols, and decisions directly within workflows, then structure that information so it can be connected across teams and projects,” she says. “The system learns from successful and unsuccessful experiments, and over time it can help design new experiments, troubleshoot problems, and recover knowledge that would otherwise disappear.” The value lies in accumulation. As more experiments are captured, the system becomes a record not just of outcomes, but of how those outcomes were reached.
That positions Labtree differently from many AI-for-science efforts. It is not trying to replace the scientist or generate conclusions from published data alone. It is trying to preserve the reasoning that sits between intention and result, and to make that reasoning available for future work.
Giving clinicians a clearer view
Illuminant Surgical operates at the point where diagnosis and intervention meet. Its platform, Skylight, is a real-time projection system that overlays medical imaging directly onto the patient, allowing clinicians to see anatomical structures in context rather than on a separate screen. The company describes this as a precision visualisation system, combining tracking, projection, and computer vision to improve spatial awareness during procedures.
“Doctors are still working with limited visibility when they operate inside the body, and every millimetre matters,” James Hu, Co-Founder of Illuminant Surgical, says. “That leads to outcomes that are difficult to accept, repeated surgeries, failed pain interventions, and misdiagnoses that could have been avoided with better visual guidance.”
The system is built around two components. “The first is SkinMatch, where we place fiducials on the body to align imaging directly onto the patient,” Hu explains. “Because we use multiple points, the system remains accurate even with movement or deformation, and we can achieve sub-millimetre precision. The second is the projection layer. We use high-definition projection and computer vision to map information onto the patient’s surface in a way that is meaningful for the clinician. The interface can be tailored depending on the procedure, because different specialists need to see different information.”
The goal is to reduce the cognitive load of translation. “Instead of looking back and forth between a screen and the patient, the information is presented directly where it is needed,” Hu says. “That improves accuracy, but it also makes these systems more accessible, because they are easier to use and integrate into existing workflows.” The company is currently completing validation studies and preparing for clinical deployment, with the aim of moving from controlled testing into real procedural environments.
Where precision becomes consequential
Across these companies, the common thread is not simply the application of AI to healthcare. It is the attempt to recover precision in systems that have become difficult to navigate, difficult to predict, or difficult to trust. Whether that is understanding the cost of care, predicting the success of a drug, preserving scientific reasoning, or improving visual guidance during procedures, the problem is the same. Information exists, but it is not usable when it matters most.
That is what makes healthcare such a demanding environment for AI. The margin for error is narrow, and the consequences of failure are immediate. A model that performs well in isolation is not enough. It has to operate inside systems that are complex, fragmented, and often resistant to change.
This marks another step in the progression of the series. The first article explored how AI is entering the physical world. The second examined how it operates within industrial systems. Here, the focus shifts to environments where precision is not optional. The next group of start-ups emerging from NVIDIA GTC moves into enterprise workflows, where the constraints are different, but the requirement remains the same, intelligence must function where people depend on it.
All companies featured in this article are part of the NVIDIA Inception programme, which supports startups developing cutting-edge technologies with access to NVIDIA’s expertise, tools and go-to-market resources. The initiative is designed to help early-stage companies scale faster and bring advanced AI-driven innovations into real-world deployment.
