Ametis 3 — Clinical Co-Pilot (vertical slice)

Real-time HUD that listens to a counselling session via the browser mic, transcribes locally with the Web Speech API, sends the text-only transcript to GPT-4o, and renders a 4-tier colour HUD (Teal → Amber → Indigo → Red) plus a just-in-time therapeutic prompt.

This is the vertical slice, not the full Ametis 3 spec. Still to build: weighted signal scoring (audio/biometric/camera), 9-stage session lifecycle, flashcard engine over the JSON theory library, auto-documentation (SOAP/DAP/BIRP/Narrative), cool-down + next-session memory loop, governance checklist.

Privacy posture

Audio never leaves the browser — only the recognised text is sent to the serverless function, which forwards it to OpenAI. To make the brief's "zero-retention audio" claim contractually true for text as well:

1. Get your OpenAI org onto Zero Data Retention (enterprise agreement).
2. Set OPENAI_ZDR=true in your Vercel env vars (or .env.local locally).
3. Confirm in GET /api/health.

Without ZDR, OpenAI may retain inputs up to 30 days for abuse monitoring. The /api/health endpoint surfaces this status honestly.

Local dev

cp .env.example .env.local      # add OPENAI_API_KEY
npm install
npx vercel dev                  # serves Vite + /api/* on http://localhost:3000

Open in Chrome or Edge (Web Speech API isn't supported in Firefox/Safari desktop). Click Start session, grant mic access, speak.

npm run dev works for the UI alone but won't serve the /api functions — use vercel dev for the full stack.

Deploy

See DEPLOY.md. One Vercel deploy hosts both the SPA and the serverless functions.

Layout

src/                   Vite + React + TS + Tailwind (UI)
  App.tsx
  components/HUD.tsx
  components/TranscriptPanel.tsx
  hooks/useSpeechRecognition.ts
  lib/signals.ts
  types.ts

api/                   Vercel serverless functions (Node.js)
  analyze.ts           POST /api/analyze
  health.ts            GET  /api/health
  _lib/analyze.ts      shared OpenAI + schema logic