How we make medical AI answer from evidence, and how we check that it did.
A doctor's question deserves an answer that comes from the right source and can be traced back to it. Our research asks how to find that source reliably, how to keep a language model's answer tied to it, and how to measure both honestly before anything reaches a clinician.
Research areas
How a clinical assistant finds the passage that answers a doctor's question across guidelines, Indian national sources, drug labels, textbooks and the medical literature, and keeps every answer tied to the exact source. We measure search stage by stage.
Which models, and which adaptations of open models, give the most reliable medical search and answers for the cost. A change is adopted only when it beats what we have on questions it was never tuned on.
Test questions built from the sources; strict scoring, where a question that needs two sources counts only when both are found; held-aside questions never used for tuning; and run-to-run variation measured, so a change smaller than the noise is never reported as progress.
Before a clinical AI feature reaches doctors it must pass tests built to break it, covering emergencies, doses and questions where guessing is dangerous, with clinician review. The system asks rather than guesses, and a safety regression blocks a release.
Papers & preprints
Papers, preprints and technical posts will be listed here as they are published. We publish a result only after it has been checked independently.
Collaborations
We welcome collaborations with clinicians and academic groups on how medical AI should be tested. Write to support@nball.ai.
AI tools for doctors in India, released one at a time.