FormDoctor：用 Claude 语音听写自动填写澳洲政府医疗表单

FormDoctor

Voice dictation to filled Australian government medical forms. Doctors dictate, identifiers are stripped, Claude extracts the fields, and the official PDF is filled for review. Open-source portfolio project — not clinically deployed.

Try it without signing up at /demo (fictional patients, cached extraction, zero paid API calls).

Live at formbridgegp.com

The Problem

Australian GPs spend hours each week manually completing government medical forms (Centrelink, WorkCover, TAC). Clinics in lower socioeconomic areas face even higher volumes. This is time that should be spent with patients.

How It Works

Doctor dictates clinical notes
        ↓
Real-time speech-to-text (Deepgram Nova-3 Medical)
        ↓
De-identify patient data (local regex, before any external API call)
        ↓
AI extracts form fields from dictation (Claude)
        ↓
Fill government PDF template (pdf-lib)
        ↓
Doctor reviews, edits, downloads

Features

• Voice dictation — Real-time transcription with medical vocabulary support
• Guided dictation — Form-specific structured prompts improve extraction accuracy
• AI form filling — Claude maps clinical notes to the correct form fields
• Privacy-first — Identifiers are stripped before the LLM call; audio is never stored. Completed forms are saved to the doctor's account; patient records only when the doctor chooses
• PDF generation — Fills official government PDF templates directly
• Doctor profiles — Set up once, auto-fill provider details on every form
• Completeness checking — Identifies missing required fields and prompts the doctor
• Review & edit — Full PDF preview with inline field editing before download

Supported Forms

Adding a new form = JSON schema + PDF template. No code changes required.

Tech Stack

• Framework — Next.js 15, React 19, TypeScript
• UI — Tailwind CSS, shadcn/ui
• Speech-to-text — Deepgram Nova-3 Medical (real-time WebSocket streaming)
• AI extraction — Anthropic Claude API
• PDF — pdf-lib
• State — Zustand (memory-only, never persisted — by design)
• Auth & audit — Supabase (PostgreSQL)
• Address autocomplete — Google Places API

Getting Started

Prerequisites

• Node.js 18+
• pnpm

Setup

# Install dependencies
pnpm install

# Copy environment variables
cp .env.example .env.local

Add your API keys to .env.local:

DEEPGRAM_API_KEY=           # Speech-to-text
ANTHROPIC_API_KEY=          # AI extraction (Claude)
NEXT_PUBLIC_SUPABASE_URL=   # Database
NEXT_PUBLIC_SUPABASE_ANON_KEY=
SUPABASE_SERVICE_ROLE_KEY=
NEXT_PUBLIC_GOOGLE_PLACES_API_KEY=  # Address autocomplete

Run

pnpm dev

Open http://localhost:3000.

Privacy & Security

• De-identification before the LLM — names, DOB, address, Medicare/CRN, phone and email are stripped locally (src/lib/deidentify.ts) before clinical notes reach Claude
• In-flight data stays in memory — transcription and patient details live in a Zustand store with no persist middleware (never localStorage)
• Completed forms are saved to the doctor's account (Supabase, row-level security per doctor) so they can be revisited; patient records are saved only when the doctor chooses
• Audio is never stored — streamed to Deepgram for live transcription, not written anywhere
• Third-party data-handling terms (Anthropic, Deepgram) are those providers' own; nothing here is a compliance claim

Project Structure

src/
├── app/
│   ├── (auth)/           # Login, register
│   ├── (dashboard)/      # Main app (form wizard, dictation, settings)
│   └── api/              # Backend routes (form processing, Deepgram tokens)
├── components/
│   ├── dictation/        # Recording, transcription, guided panel
│   ├── forms/            # Form selection, patient details, review
│   └── ui/               # shadcn/ui components
├── lib/
│   ├── schemas/          # Form schemas (JSON) + PDF templates
│   ├── stores/           # Zustand state management
│   ├── deidentify.ts     # PII stripping (local regex)
│   ├── reidentify.ts     # Merge patient data back after LLM
│   ├── llm.ts            # Claude API client
│   ├── pdf-filler.ts     # PDF generation
│   └── ...
└── types/

License

MIT