Internal Medicine · West Bloomfield, Michigan
Board-certified internist and hospitalist. I build clinical AI that has to earn its answers: model output is treated as a proposal and accepted only against evidence outside the model.
Measured results from a fail-closed clinical knowledge engine: assertion accuracy on held-out notes, concept-linking precision, and how frontier models behave as judges and readers of clinical text.
How it works · Trials · History · Lessons · Testing
On the NOHARM clinical-safety benchmark, one structured instruction moved five frontier models' scores more than the choice between them, and the plan format, not length, carried the gain.
Data and code: github.com/OJamals/noharm-prompting