BioStack Raises $9.5M Seed for Medical AI Training Data

A hospital record becomes more useful to an AI developer when it preserves what happened after a treatment decision. BioStack is building a business around keeping that sequence intact, connecting clinical information to the training and evaluation of healthcare AI.

The San Francisco company announced $9.5M in seed funding on October 22, 2026. Co-founders Sanat Mishra, CEO, and Parth Patwa, CTO, are developing infrastructure that gives model builders patient context and a place to practice using it. The financing puts resources behind a demanding supplier assignment: prepare the clinical material, shape a useful exercise and help determine whether the model learned anything worth keeping.

What BioStack's $9.5M seed will fund

BioStack's release names Peak XV, Afore Capital, DCVC, Emerson Collective, SV Angel, Verdict Capital, Streamlined Ventures and Y Combinator as backers. The company says the round will support expansion in the US and internationally, technology-platform development, and engineering and research hiring. It has not supplied a valuation or a lifetime capital total in that announcement.

Those spending priorities fit a company whose work runs between healthcare institutions and model developers. Adding a source of clinical information can create work in collection, preparation and interpretation before a researcher has an exercise to run. Geographic expansion therefore reaches into the product itself: different records have to become useful inputs to a clearly defined research task.

For a buyer, the useful question is how much of that preparation the supplier takes responsibility for. A lab purchasing data may still have to establish which events belong to one history, what the model should know at each point and which outcome can fairly judge its decision. A supplier that delivers more of that work has a larger role in the customer's research process. That role also gives the buyer more to inspect when deciding whether to use the resulting environment. DevCuration's coverage of clinical-data infrastructure offers an adjacent part of this research-and-care data discussion.

How longitudinal data becomes a training product

BioStack describes its data offering as connected layers of a patient's history, including clinical records, imaging, laboratory information, treatments and outcomes. The significance lies in the relationship between events. A scan, a prescription and a later result can each be accurate while still being insufficient to reconstruct the information available at an earlier decision.

That timing matters for a training exercise. If a model can see the later result while being asked to make the earlier choice, the exercise may reward information it would never have possessed at the relevant moment. The supplier's preparation choices therefore help define what the exercise measures. This is an analytical consequence of longitudinal training, rather than a claim about a particular customer's deployment.

BioStack reports a network spanning over 100M de-identified patient records across the US, India, Brazil and Argentina. That is a company-reported measure of network reach. A researcher evaluating an individual cohort would still need to understand its completeness, available modalities and suitability for the question being asked.

The commercial unit becomes more specific as the research question sharpens. An AI lab may need a sequence with a useful follow-up period; another buyer may need an evaluation task built around a particular kind of observation. The value of the service depends on how well the supplied material supports that defined job, which is something a headline record count leaves open.

Why the scoring system belongs in the purchase

The company's Y Combinator profile describes workflow-aligned environments for post-training and reinforcement learning. In that setting, the supplier helps construct both the situation a model encounters and the feedback it receives. Decisions about the grading process become part of the product a customer is evaluating.

BioStack's founders have made that second assignment visible in their own research on medical AI graders. The company-authored study describes testing whether automated graders notice changes that alter a medical answer's meaning. Its reported failures illustrate a concern about evaluation design; they are not an independent clinical validation of BioStack's platform or a general failure rate for deployed healthcare systems.

The business implication is practical. A lab needs to know why a model received a favorable score and what would cause that score to change. A grading method that overlooks a consequential change can make a training run appear more successful than the underlying task warrants. Reviewing the grader becomes part of reviewing the service.

For BioStack, that opens a technical conversation beyond delivering a collection of files. Customers can ask about case construction, evaluation criteria, the evidence available to a grader and the role of expert review. Each answer affects how confidently a research team can interpret its results and how much additional validation it must perform in its own setting.

Two kinds of expertise behind the expansion

BioStack's leadership page places complementary backgrounds beside this product. Sanat Mishra's stated experience includes cancer genomics and computational biology. Parth Patwa previously worked in generative AI at AWS and in research at MIT. The founders are bringing knowledge of biological data and AI systems into the same supplier relationship.

The company's open positions make the division of work particularly clear: a clinical data lead and a founding RL/reasoning research engineer. Those are different assignments with a shared dependency. The clinical information has to support an appropriate task, while the training system has to provide useful feedback about a model's performance on that task.

The $9.5M round can support hiring and platform development across that handoff. BioStack has not disclosed enough in the announcement to establish current revenue, named hospital partners or independently measured customer gains. Buyers can instead examine the work that is visible: the described data products, the founders' research and the roles the company is recruiting to fill.

As BioStack expands its network, the clinical data lead and the reinforcement-learning engineer will keep meeting inside the same patient case. One has to preserve the history that makes the exercise meaningful; the other has to make its feedback useful to the team building the model. The financing gives the founders more room to keep those two jobs connected as new records and new research questions arrive.