Engineer builds AI medical training simulator by strictly separating language from clinical facts

A backend engineer and his doctor co-founder built Rounds, an AI-powered patient simulator designed to help medical students practice clinical reasoning beyond what standard exams offer. Early prototypes revealed a critical flaw: the language model would alter symptom histories, invent lab values, and give away diagnoses when prompted cleverly. To fix this, the team separated clinical truth — stored in a fixed case state linked to a medical knowledge graph — from the language model, which only controls how the patient communicates, not what the facts are. Investigations and examination findings are retrieved from authored case data and cached per session, so repeated queries always return the same result. Grading works similarly, using an evidence log of student actions checked against deterministic rules before any AI-interpreted credit is awarded.