Archangel Health, a two-person San Francisco startup in Y Combinator's Fall 2026 batch, builds reinforcement learning environments inside de-identified electronic health records to train and test medical AI. No public source lists a funding round for the company.
Its wager is that regulators and hospitals, who increasingly want evidence that clinical AI is safe before it goes live, will not accept exam-style questions as that evidence. Proof, in Archangel's view, has to come from watching a model work inside the record itself.
What the company says it has built
Y Combinator's listing, updated in November 2026, describes the company as building "the infrastructure to post-train and evaluate medical AI." In practice, that means reinforcement learning environments set inside real de-identified records. It also means training data verified by physicians, plus benchmarks meant to probe clinical reasoning on hard cases.
The listing says frontier labs and medical AI companies use the product today. It names none of them, and the public sources reviewed (the YC listing and the company's product page) name no customers or pilot sites. The stated next step is to post-train open-source models for health systems, using each system's own data and its own physicians' judgment.
A third-party GitHub roadmap document, with a commit dated October 27, 2026, fills in some detail. It describes sandboxed EHR tool use with step-by-step grading, multi-visit tasks scored against what happened later, physician-written training data and rubrics, and reinforcement learning driven by verifiers (automated checks of whether a model's answer is correct). The same document lists a Medical Guideline Benchmark as "coming soon." No public release has surfaced.
The founders
Y Combinator lists Aryaa Bhatia and Tej Patel as founders. Patel was previously chief of staff at SimCare AI (YC S24) and an AI fellow at Define Ventures. He also led machine learning research at Cedars-Sinai, where he was first author on a patent-pending graph neural network for predicting hematomas, and he holds a bachelor's degree in neuroscience and philosophy from UC Berkeley.
Bhatia co-founded Plureto, which ran tactile-sensing robotics pilots in hospitals and on farms. Her YC profile attaches a $2.5 million valuation to that company. She has also done research at Lawrence Livermore National Laboratory, and she studies industrial engineering and operations research alongside electrical engineering and computer sciences at Berkeley.
A data provider profile on Opendatabay, which joined the marketplace on August 24, 2026, lists the legal entity as Amma Healthtech Inc. It describes physician-graded, consented, de-identified datasets that come with datasheets and provenance records.
Why static tests fall short
The pitch tracks a real shift in how researchers judge medical AI. A paper in BMJ Digital Health from October 2026, "State of clinical AI in 2026," found that knowledge benchmarks are saturated, while performance on agentic tasks inside EHRs remains lower. It proposes a four-level evaluation ladder that ends in prospective trials.
A 2025 JMIR review of 39 benchmarks landed in similar territory. Top models scored in the 70-79% range on multiple-choice knowledge tests but did worse on open, practice-based tasks. A PLOS Digital Health paper published May 26, 2026, cautioned that gains on benchmarks may not translate into clinical usefulness.
Academic groups are building comparable sandboxes: MedAgentBench and FHIR-AgentEval among them. EHR-Complex draws on MIMIC-IV and spans 52,000 tasks. A July 2026 arXiv paper on reinforcement learning in FHIR environments identified what it called a "silent-finish ceiling" in early designs. Microsoft's HealthAgentBench, for its part, has retired tasks that became too easy.
Pressure from regulators and clinicians
ECRI, the patient safety organization, ranked "Navigating the AI Diagnostic Dilemma" as the top patient safety concern for 2026 in a report dated March 9, 2026.
Federal money is moving in the same direction. ARPA-H announced its ADVOCATE program on September 9, 2026, with the aim of producing what the agency calls the first FDA-authorized clinical agentic AI for cardiovascular care. Kaiser Permanente will embed agents in Epic workflows across 21 medical centers and more than 260 clinics, using shadow-mode deployments (where the AI runs quietly alongside clinicians without influencing care) and pragmatic randomized trials. Duke will run multi-site validation.
Rick Abramson, M.D., director of the FDA's Digital Health Center of Excellence, said in the release that the program "is the kind of ambitious effort that will be critical as we think about continuous ongoing monitoring of autonomous clinical AI for both safety and effectiveness."
The American Medical Association adopted policy on June 10, 2026, calling for transparency and physician oversight of AI used in clinical decision support. "There must always be transparency, accountability, and meaningful physician oversight," said John Whyte, M.D., M.P.H., the AMA's chief executive.
Abroad, the timetable is staggered. Under the EU AI Act, high-risk rules for AI embedded in products, medical devices included, apply from August 2, 2028, according to the EU's implementation timeline as adjusted in 2026. In the United States, the algorithm transparency provisions of the HTI-1 rule face a 2025 proposal to strip them out. The Federal Register draft states: "we propose to remove transparency and risk management requirements established in the HTI-1 Final Rule." No final decision turned up in the sources reviewed.
Competitors and the market
The sources reviewed cite no direct competitors to Archangel. The nearest comparisons are academic and big-tech benchmarks, such as OpenAI's HealthBench, which contains 5,000 physician-written conversations with grading rubrics. Another is PhysicianBench, a physician-graded tool-calling benchmark built around 14 FHIR tools. Archangel's products page references both, along with MedHELM and MedAgentBench.
The customers it would serve are flush, or at least growing. TechCrunch reported that OpenEvidence raised a $250 million Series D at a $12 billion valuation on January 21, 2026. Rock Health data summarized by TechTarget put U.S. digital health funding at $7.4 billion in the first half of 2026, driven largely by AI. FierceHealthcare reported on August 17, 2026, that Abridge had moved from ambient scribing into decision support through a tie-up with UpToDate.
Adoption is uneven. An Emory study released February 4, 2026, found that 62% of U.S. hospitals using Epic had adopted ambient AI documentation by 2025. A February 2026 survey from Eliciting Insights put adoption of AI clinical decision support at 28%, against 68% for ambient listening.
What to watch
Archangel's claims rest mostly on its own listing and a handful of third-party references, which is thin footing for a company asking hospitals to trust its yardstick. The Medical Guideline Benchmark would be the first public test of whether its hard-case evaluations hold up. The company has not named a customer, and no funding round has been made public.
For health systems, the YC listing says the next phase is post-training open-source models on their own data and their physicians' judgment.
