Form Reader

Snap a form, get clean structured data back.

A thin AI extraction layer for handwritten forms and general documents. Users scan with whatever they already trust — iOS Notes, the Files app, Google Drive, Adobe Scan — and upload the result here. The tool extracts fields into structured JSON, shows a side-by-side review, and exports the cleaned data wherever it needs to go.

Built for The Davis Financial Group as their first-client intake tool, but designed to be generic enough that the same pipeline handles any form (or any document) with a different prompt.

Why this exists

Every serious document-scanning app (iOS Notes, Google Drive, Adobe Scan, Microsoft Lens) already has excellent native camera capture, edge detection, multi-page support, and review UX. Building a scanner from scratch in the browser would reinvent work those teams already did better than we ever could.

What doesn't exist is a simple upload-and-extract tool that knows specific forms deeply and returns clean structured data the advisor can actually use. So we built that layer and nothing more.

What it does

• Accepts images (JPG / PNG / HEIC) and PDFs — single or multi-page, any combination
• Sends them to a vision model with a domain-specific prompt
• Returns structured JSON with per-field confidence scores
• Shows a side-by-side review: original pages on the left, extracted fields on the right
• Amber-flags anything the model was unsure about
• Lets the user edit any field inline before exporting
• Exports via email (mailto), clipboard, or JSON download

What it doesn't do

• No camera UX of its own — uses the browser's file input, drag-and-drop, paste, and the Web Share Target API (PWA)
• No data persistence — images stay in the browser, results stay with the user
• No user accounts, no history, no database
• No client-facing outputs — this is a back-office tool

Modes

Davis mode (default). Knows the three Davis Financial intake documents cold (Quick Start, Financial Sketch, What to Bring) and extracts into their schema. Multi-page forms are sent as one request.

General mode ("digitize anything"). Takes any photo or PDF and returns clean markdown, a one-line summary, and extracted key facts. For notes, receipts, letters, whiteboards, anything.

Running locally

git clone https://github.com/UBIpromoter/form-reader.git
cd form-reader
npm install
export OPENROUTER_API_KEY=sk-or-...
node server.js

Then open http://localhost:8091.

Configuration

Environment variables:

• OPENROUTER_API_KEY (required) — OpenRouter API key
• PORT (optional, default 8091)
• MODEL (optional, default google/gemini-2.0-flash-001) — any OpenRouter vision model
• KILL_SWITCH=true — returns 503 on all extraction requests. For shutting off a public instance fast.

Abuse protection

For public deployments:

• 20 requests per minute per IP (in-memory rate limit)
• 15 MB maximum payload per request
• 10 files maximum per request
• Kill switch via KILL_SWITCH=true env var

Architecture

• Server: 300 lines of Node. No frameworks. Just Busboy for multipart parsing, standard https for the OpenRouter call.
• Frontend: single HTML file, no build step, no dependencies. Vanilla JS + CSS.
• Model: Gemini 2.0 Flash via OpenRouter. Cheap, fast, good at vision.
• Prompt: domain-specific Markdown files in prompts/. Swap the prompt, swap the domain.

Adapting to a different form

1. Copy prompts/davis.md to prompts/<yourform>.md
2. Edit the field list and output schema to match your form
3. In server.js, add the new prompt to the PROMPTS map
4. In public/index.html, add your field metadata to DAVIS_FIELDS (rename or fork)

Or fork the repo and replace both entirely.

Hosting notes

Works on any VPS that can run Node. Behind a reverse proxy for TLS. For the Davis demo it runs on a home server with a Cloudflare tunnel; for real deployment it would live on whatever infrastructure the client controls.

License

MIT.

Credits

Built by Green Wall AI Consulting as the first-client deliverable for The Davis Financial Group. The scanner work is deliberately not reinvented — we lean on iOS Notes, Files, Google Drive, and other native scanners for capture, and focus our effort on the AI extraction layer.