About Osseus
We are building
medical superintelligence.
Our mission is to make medical AI work reliably across the populations and places it will serve, including patients left out of today's training data.
Why we exist
Medical AI must work for everyone it serves.
Medical AI is improving quickly, but average scores can hide differences between patient groups. Our health AI research found that frontier models performed worse for minority groups on the tasks we studied. A system can look strong overall and still be less reliable for some patients. In medicine, that is not a secondary fairness issue. It is a basic failure of trustworthiness.
Performance also remains poor when models have to complete whole workflows. As of September 2026, the leading published systems finish 55% of HealthAgentBench tasks, 50% of PhysicianBench tasks and 36.3% of HealthAdminBench tasks. We track these gaps on our medical and biological leaderboard.
These gaps are serious, but they are not inevitable. We can measure performance by demographic group and care setting. We can train models on data that better reflects the people they will serve, then test them with clinicians before they take on more responsibility. Healthcare produces histories, decisions, and outcomes that can support better training and evaluation. The hard work is collecting them with consent, governing them carefully, and making sure the resulting datasets represent the people a model will serve.
Osseus brings those pieces together. We work with hospitals and clinics around the world to build representative clinical datasets and realistic training environments. We use that foundation to measure performance across patient groups, reduce gaps where we find them, and bring medical AI into care with clinicians in control. That is our mission: medical AI that clinicians can trust to work for more of the people they treat.
Why Osseus
Clinical AI is trained on too narrow a range of patients.
Most clinical AI is trained on data from a handful of countries, leaving many patient populations underrepresented. When models are used beyond those populations, accuracy can fall. That makes them less fair, and less effective.
Our partnerships help close that gap: better medical AI, built with and for more of the world.
More than half of the datasets behind published clinical AI come from two countries, the United States and China. Almost every one of the most-used databases is from a high-income country.
PLOS Digital Health, 2022The imbalance is in who is enrolled as much as where. Of the participants behind genome-wide association studies, 86.5% are of European descent, while those recorded as African, not counting African American or Afro-Caribbean samples, are 0.47%.
Cell Genomics, 2024Of 106,950 publicly available skin cancer images, ethnicity was recorded for 1.3% of them. Among those, not one patient was of African, Afro-Caribbean or South Asian background.
The Lancet Digital Health, 2022The models inherit it. Chest X-ray classifiers trained on three of the largest public datasets consistently underdiagnosed Black, Hispanic and female patients, telling them they were healthy when they were not, and were worst of all at the intersections.
Nature Medicine, 2021It costs accuracy, not only equity. Risk scores built from European cohorts keep 51% of their predictive power when applied to South Asian patients, 47% for East Asian and 39% for African.
Human Genomics, 2024
Careers
Join us to build at the most important frontier of AI, human health.
- Research scientists
- Research engineers
- Clinician-builders
- Deployment engineers
- Product engineers
- Not on this listWrite anyway and tell us what you would bring.
