Scribe

Scribe captures a consultation from the browser microphone, transcribes it with OpenAI Whisper, and turns the transcript into a structured clinical note using templates parsed from a clinician's own export. Pick a template, record, stop, review the transcript, and generate the note.

Because a clinical note that invents facts is worse than none, a second verification pass checks every claim in the generated note against the transcript — an anti-hallucination guard built into the workflow, not bolted on after.

mic (browser) ──▶ /api/transcribe ──▶ Whisper ──▶ transcript
transcript + template ──▶ /api/note ──▶ chat model ──▶ clinical note
                                                          │
                                    /api/verify ──▶ fact-check each claim vs transcript

Templates are structured on the fly: scripts/parseTemplates.js parses a garbled EMIS-style export into clean templates.json. The note prompt is built to resist hallucination, and /api/scribe runs transcribe → generate → verify in a single call.

• Mic-to-note in the browser

  In-browser MediaRecorder capture, a 25 MB Whisper upload cap, and a review step before anything is generated.

• Template-driven structure

  Notes follow real clinical templates — history, examination, assessment and plan — parsed from the clinician's own export into clean JSON.

• Verification pass

  Every claim in the note is checked against the transcript in a second pass, so the output is grounded in what was actually said.

• One-call pipeline

  /api/scribe takes audio plus a template ID and returns transcript, note, and verification together.

Backend	Node.js · Express API + static UI (server.js)
Transcription	OpenAI Whisper (lib/transcribe.js)
Note gen	Chat model, anti-hallucination prompt (lib/generateNote.js)
Verify	Claim-by-claim fact-check (lib/verifyNote.js)
Templates	Parsed from export → templates.json
Frontend	Browser UI with MediaRecorder mic capture