Medical AI Scribe Safety

A scribe that extracts facts from a spoken consult can be confidently wrong, and a silent error in a clinical note is dangerous. The output had to be verifiable: medication mistakes surfaced rather than quietly accepted, note fields that do not claim more certainty than the conversation supports, and edit paths that cannot quietly rewrite the surrounding text.

• In production: a misheard but real medication is surfaced for review instead of being silently accepted
• Clinical note fields expose status and confidence, so uncertain lines stand out instead of reading as fact
• Debugged a corrupted-recording failure down to the byte level (two overlapping recorders inserting a chunk before the WebM header) and fixed it with a re-entrancy guard

• Deterministic medication terminology validation: spoken drug names are matched against a validated coding rather than accepted as a raw LLM guess
• Medication review with click-to-fix: each recognised drug term gets a colour-coded status (including exact matches previously accepted silently) and is one-click correctable, with a training-consent step defaulting to off
• Fact-grounded WSOEP note fields: status and confidence are derived purely from the extracted facts, values are nulled when the facts don't support them, and human review is forced on non-present facts
• Note edits reject blank input and insert the chosen name literally instead of via a regex, closing a path that could silently corrupt the record
• Teams remote-consult recording: tab and screen audio is mixed with the doctor's microphone via the Web Audio API into the existing transcription pipeline, feature-flagged and cross-platform
• Verified with TDD (pytest and Vitest) plus adversarial multi-agent code audits that found and fixed real bugs, including stream leaks and 7 medication-review edge cases