San Francisco-based BioStack has raised $9.5 million in seed funding to build a global clinical data infrastructure platform for training and evaluating healthcare AI models.
The financing includes Peak XV Partners, Afore Capital, DCVC, Emerson Collective, SV Angel, Verdict Capital, Streamlined Ventures and Y Combinator.
Founded in 2025 by CEO Sanat Mishra and CTO Parth Patwa, BioStack participated in Y Combinator's Spring 2026 batch.
Mishra previously worked on cancer genomics research at institutions including Stanford, Yale, Carnegie Mellon and the Max Planck Institute. Patwa previously worked on generative AI research at Amazon Web Services and conducted research at MIT.
The company is addressing one of the most significant constraints in medical AI development: access to high-quality clinical data.
Healthcare organizations generate enormous amounts of information, but the records are frequently fragmented across hospitals, laboratories, imaging systems and electronic health record platforms.
A patient's medical history may include clinical notes, blood tests, imaging, medications, treatment decisions and outcomes collected over several years.
These records are often stored in incompatible formats or disconnected systems, making them difficult to use for machine learning.
BioStack is building infrastructure that connects and structures this information into longitudinal datasets.
The company reports access to a network spanning more than 100 million de-identified patient records across the United States, India, Brazil and Argentina.
Its system combines clinical notes, laboratory results, diagnostic images, treatment histories and outcomes to reconstruct patient journeys over time.
The company says it has already integrated several hundred thousand medical images, including X-rays, MRIs and CT scans, alongside millions of telehealth audio transcripts and other clinical information.
These data sources create opportunities beyond traditional medical AI training.
Many current medical models are evaluated primarily on medical examinations, question-answering benchmarks and relatively structured clinical datasets.
Real clinical practice is more complicated.
Patient information arrives over time, tests may be incomplete, symptoms change, and treatment decisions must account for uncertainty.
BioStack aims to reproduce those conditions through realistic training and evaluation environments.
Its platform can convert clinical histories into tasks that require models to interpret available information, make decisions and receive feedback based on subsequent observations.
For example, an AI model could be evaluated on its ability to identify a likely diagnosis, determine an appropriate next test or interpret changes in a patient's condition.
The company also intends to support reinforcement learning, allowing developers to improve models using feedback from clinically grounded tasks.
This positions BioStack as an infrastructure provider for AI laboratories, biotechnology companies and healthcare researchers rather than a consumer-facing medical application.
Its potential customers need access to representative clinical data, evaluation tools and realistic environments for improving healthcare AI systems.
BioStack also supports applications in medical research, biomarker discovery, patient stratification and treatment-response analysis.
In India, the company says it is helping connect eligible cancer patients with an investigational cancer vaccine program, while supporting the use of clinical information for related biotechnology research.
The new financing will fund expansion of its data network, technology development and hiring across engineering and research.
The business opportunity is substantial because medical AI systems require more than strong general-purpose language models.
Clinical performance depends on access to reliable information reflecting real patients, diverse healthcare environments and outcomes over time.
BioStack's geographic coverage could provide a competitive advantage by supporting datasets spanning different populations and healthcare systems.
However, the scale of its record network should not be confused with the number of fully standardized, validated or immediately usable training examples.
A record count alone does not establish data completeness, representativeness, linkage quality or clinical relevance.
Healthcare data also creates serious privacy, security and governance requirements.
De-identification reduces privacy risk but does not eliminate every possibility of re-identification, particularly when different datasets are combined.
BioStack must demonstrate that its data collection and use comply with the applicable legal requirements, institutional agreements and consent or authorization frameworks in each market.
The company also needs to establish that its training environments produce measurable improvements in model performance on independent clinical evaluations.
The most defensible part of its business may ultimately be the difficulty of building and maintaining high-quality data relationships rather than simply the volume of records collected.
Its next challenge is turning its reported data access into validated products that AI developers and research organizations are willing to purchase repeatedly.
If successful, BioStack could become part of the underlying data and evaluation infrastructure used to develop future medical AI systems.