Health data & AI infrastructure · Training environments for medical AI
BioStack Platforms
A San Francisco startup that turns real, messy clinical and preclinical records into environments where AI labs can train models to practice medicine, not just recite it. Founded in October 2025, it has one investment on the public record and no named customers yet.
The call that changed the pitch
BioStack was not pitching itself when the idea that stuck finally showed up. The founders were on a sales call, trying to sell drug discovery data to one of the largest human data companies in the world. Halfway through, the buyer changed the subject. According to the company's own account of the meeting, the buyer did not want another generic data vendor. It asked whether BioStack could deliver high quality clinical data under a six figure contract.
“That call changed the company,” the founders wrote in their Y Combinator launch post. What followed was a reframing: the real bottleneck for healthcare AI was not raw data alone, but the full loop a model needs to get better at clinical work, meaning data, simulation environments, evaluations and reward functions.
The problem BioStack says it is solving
The company's pitch is that healthcare AI is trained on a version of medicine that is cleaner than the real thing. “Real patients are messy. Their records are incomplete. Their labs are late. Their notes conflict,” the founders wrote. “Doctors do not just answer questions. They make decisions with imperfect information, watch what happens, and adjust. That is the loop AI needs to perfect.”
BioStack says today's AI labs already have medical data but lack real clinical environments where a model can act, get scored, and improve. Its stated approach is to convert longitudinal patient records, notes, labs, imaging, medications, vitals, diagnoses and outcomes, into tasks a model has to work through: diagnose this, order the next test, adjust the medication, flag the risk, predict what happens next. The model's move is then scored against outcomes, documented reasoning, clinical guidelines and, the company says, clinician review. None of that scoring pipeline or its accuracy has been independently examined; it is described only in the company's own materials.
Two researchers, not operators
Sanat Mishra, the CEO, describes himself as a former cancer genomics researcher at Stanford, Yale, Carnegie Mellon and the Max Planck Institute, and as an early researcher with top AI labs on reinforcement learning tasks and benchmarks in healthcare and biotech. He is from Mumbai and holds a BS from IISER Mohali and an MS from Carnegie Mellon, according to his Y Combinator founder bio. Independent confirmation is partial but real: a CDK11 cancer biology preprint posted to bioRxiv in August 2025 lists a Sanat Mishra as a co-author alongside researchers at Yale University and Cold Spring Harbor Laboratory, consistent with his stated genomics research background, though the preprint does not identify him as BioStack's founder.
Parth Patwa, the CTO, says he was a generative AI scientist at AWS and a researcher at MIT before that, with more than 40 papers and 1,800-plus citations across venues including CVPR, NeurIPS and EMNLP. His own academic homepage and an Amazon Science author page independently confirm a real publication record in natural language processing and machine learning, an MS in electrical and computer engineering from UCLA, earlier research work at the PathCheck Foundation under Dr. Ramesh Raskar, and a role at Amazon that Patwa's own site dates to February 2024 as a data scientist, not the AWS title BioStack uses on its team page. The exact paper and citation counts were not independently verified.
Two false starts before this one
The founders say BioStack tried two other ideas first. The first was a multi-agent system for biology research, killed after two months because, in their words, lab workflows are messy and fragile enough that “one wrong parameter compounds downstream and suddenly nobody trusts the output,” and because scientists were not ready to hand real work to agents that still felt brittle.
The second was generating preclinical drug discovery data itself, through the company's own lab facility. The founders describe that version as capital intensive, slow and operationally painful, with an edge that amounted to cheaper offshore execution rather than a real moat. They killed that one too, before the sales call redirected the company toward simulation environments and post-training infrastructure for AI labs.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Robinhood Ventures Fund II holding | A schedule of investments in a 424B1 filed Aug 13, 2026 lists BioStack Platforms, Inc., 680 2nd Street, San Francisco, CA 94107, business described as “Real-world training environments for healthcare AI,” a SAFE dated Jun 2, 2026, cost and fair value both $250,000. An earlier N-2 filed Jun 30, 2026 lists the same $250,000 SAFE but describes the business as “Data sets for training models in healthcare and drug discovery.” | Public record |
| Own SEC filings | No CIK and no Form D found under BioStack Platforms or BioStack in SEC EDGAR company search as of Sep 25, 2026. Section 4(a)(2) or other exempt raises would not show up in Form D search, so the absence does not mean no other money has moved. | Not found |
| Customer traction | The company says it has grown “from one customer conversation into 17 customers and prospects, including top AI labs,” and is seeing demand across data, evaluations and RL environments. No customer is named. | Company-stated |
| Team size | 10 people, per the Y Combinator company page. | Company-stated |
| Third party assessment | An independent YC-batch review site wrote that BioStack has “two technically credible founders with domain-relevant research backgrounds,” but also that there is no evidence yet of product traction, customers or revenue, and flagged competition from Tempus, Flatiron and Scale AI. | Independent |
| Founders' research records | Partially confirmed outside company channels: a 2025 bioRxiv cancer biology preprint co-authored by a Sanat Mishra with Yale and Cold Spring Harbor researchers, and independent academic and Amazon Science pages for Parth Patwa. Exact paper counts, citation counts and full prior job titles are not independently verified. | Independent |
Read plainly: the clearest fact on the public record is that a Robinhood-sponsored venture fund put $250,000 into BioStack in June 2026 at cost. Everything about customers, pipeline and the size of the AI lab relationships is, for now, only what the company says about itself.
What to watch
- Whether BioStack names any of its stated 17 customers or prospects, or discloses revenue.
- A Form D or other SEC filing under BioStack's own name, which would show whether it has raised a priced or SAFE round beyond the Robinhood Ventures Fund II allocation.
- Independent evaluation of BioStack's scoring pipeline against clinician judgment, which has not been published anywhere BioStack does not control.
- How BioStack differentiates itself from larger, well-capitalized incumbents already selling healthcare data and evaluation infrastructure to AI labs.
In their words
“The bottleneck is not models. It is access to high-quality biological data.”
BioStack Platforms, Y Combinator company profile · Company-stated
“Real patients are messy. Their records are incomplete. Their labs are late. Their notes conflict.”
BioStack Platforms, Launch YC post, 2026 · Company-stated
“Doctors do not just answer questions. They make decisions with imperfect information, watch what happens, and adjust. That is the loop AI needs to perfect.”
BioStack Platforms, Launch YC post, 2026 · Company-stated
“One wrong parameter compounds downstream and suddenly nobody trusts the output.”
BioStack Platforms, on its first, abandoned multi-agent biology idea · Company-stated
“That call changed the company.”
BioStack Platforms, Launch YC post, on the sales call that redirected the business · Company-stated
“We did not arrive here through a brainstorm but got dragged here by the market.”
BioStack Platforms, Launch YC post, 2026 · Company-stated
“The market is real and large, but the competition from well-capitalized incumbents (Tempus, Flatiron, Scale AI) is serious.”
The YC Tier List, review of BioStack Platforms · Independent
“This is a strong idea that needs proof of pipeline deals, LOIs, or at minimum beta partners before it warrants serious conviction.”
The YC Tier List, review of BioStack Platforms · Independent
Related companies
Sources
- Public recordRobinhood Ventures Fund II, Form 424B1, schedule of investments
- Public recordRobinhood Ventures Fund II, Form N-2, schedule of investments
- Public recordSEC EDGAR company search, “biostack”
- Public recordSEC EDGAR full text search, “BioStack Platforms” and “Sanat Mishra”
- IndependentOn-target toxicity limits the efficacy of CDK11 inhibition against cancers with 1p36 deletions
- IndependentParth Patwa, personal academic homepage and publication list
- IndependentParth Patwa author page
- IndependentBioStack Platforms review
- CompanyBioStack Platforms: Real world training envs for healthcare AI models
- CompanyLaunch YC: BioStack Platforms, Realistic healthcare simulation environments
- CompanyBioStack homepage
- CompanySanat Mishra, CEO, BioStack Platforms, profile
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
