Splitty 🧾

Scan a restaurant bill. Tap who had what. Everyone pays their share over UPI.

Six people finish dinner, one person paid, and nobody wants to do arithmetic. Splitty is a web app with no backend state for exactly that moment: photograph the bill, get structured line items, assign them with taps, split the taxes correctly (liquor VAT only hits the drinkers), and settle with a UPI QR per person.

Home	Assigning	Assign by chat	Charges	Settle up

On desktop the same flow spreads into two panes, receipt on the left with people, totals, and actions in a sticky aside:

Why it's different

It never dead-ends. Scanning tries your own Gemini key first (straight from the browser to Google), then a shared free-tier proxy, then on-device OCR with tesseract.js WASM. If all three are down you can still type the bill in by hand.

Assignment is the part competitors get wrong, so most of the work went there. You select whoever shared a thing in the bottom tray and tap the thing: it splits evenly among them in one tap. Long-pressing an item opens exact portions with half-unit steppers, for nights when someone had two and two people shared one. Tapping a name badge takes that person off an item, swiping left clears it, and a single-level undo covers every change. When a tap can't do anything, the row shakes and the phone buzzes instead of silently ignoring you.

If tapping is too much work, tap Chat and type "Asha ate the biryani, Ben and Chitra shared the naan, everyone split the fries". The AI replies with a per-item proposal card, and nothing is applied until you confirm it. This works with your own key or through the shared server tier.

The money always reconciles. Every division runs through one largest-remainder allocator, so per-person totals sum to the bill total to the paisa. An invariant check and a 1000-run fuzz test enforce it, and amounts are integer paise everywhere; floats never touch money.

Indian bills are handled properly: CGST and SGST collapse into a single GST decision, liquor VAT is scoped to the people who drank, and the service charge carries a note that it isn't mandatory (CCPA guidelines, 2022) next to a one-tap remove button.

Privacy claims here are literal, not aspirational. The server stores nothing except a shared split with a 30-day TTL. No images, no API keys, no UPI IDs. Payment details travel in URL fragments, which never leave the browser, and an "always on-device" toggle keeps photos on your phone entirely.

Settling up

• Tap a person and a full-screen QR appears with their exact amount, ready to scan from any UPI app. Android also gets a direct upi:// intent button; iOS gets per-app buttons for GPay, PhonePe, and Paytm.
• One QR or link shares the whole split with the table, either as a 6-char code backed by Vercel KV or as a fully serverless lz-string URL fragment. Paid status syncs live in both directions, and sharing again after edits updates the stored split under the same code (only the creator's device holds the write key).
• Exports go out as a receipt-styled PNG drawn on a canvas, per person with share fractions and the share link printed on it, straight to WhatsApp through the native share sheet. A clean plain-text version covers everything else.
• Recent splits stay on the device: live through their code for its 30-day life, then from the local snapshot. The app installs as a PWA on iOS, Android, and desktop.

Stack

React + Vite + TypeScript · Tailwind v4 · zustand · motion · vaul · sonner · NumberFlow · lucide-react · input-otp · react-qr-code · lz-string · tesseract.js · Vercel serverless (Edge) + Upstash Redis.

Getting started

npm install
npm run dev        # app at localhost:5173 — manual entry, BYOK scanning, on-device OCR,
                   # and fragment sharing all work with no backend at all
npm test           # split-engine suite incl. the paisa-exactness fuzz test
npm run build      # type-check + production build

The serverless pieces in api/ only run when deployed (or under vercel dev): shared-key scanning and chat with rate limiting, plus code-based shares.

Deploying to Vercel

1. Import the repo (framework: Vite). vercel.json carries the SPA rewrites and CSP headers.

2. Storage → Create Database → Upstash for Redis, then connect it to the project. This injects KV_REST_API_URL and KV_REST_API_TOKEN for you.

3. Add the remaining environment variables (see .env.example):

   Variable	Purpose
   SHARED_GEMINI_KEY	Free-tier Google AI Studio key. Use a project with billing disabled, so worst-case abuse costs quota, not money
   SHARED_GEMINI_MODEL	Optional; defaults to gemini-3.6-flash
   DISABLE_RATE_LIMIT	true only for local testing

4. Deploy. The scan proxy is rate-limited to 3 requests per IP per minute, and KV stores only rate-limit counters and computed splits.

Project layout

src/lib/        split engine (allocate, computeSplit), money, UPI links, unit helpers
src/scan/       preprocess → engine router → gemini / tesseract + geometric layout parser
src/store/      zustand stores (bill, settings, history)
src/screens/    Home · People · Items · Assign · Charges · Summary · Settings · Join · Shared
src/share/      lz-string codec, PNG/text export
api/            Edge functions: /api/scan + /api/assign (rate-limited proxies), /api/split (+/:code)

Tests run with npx vitest run and cover the engine allocation cases, the geometric parser, and the share codec round-trip. API functions type-check via npx tsc -p tsconfig.api.json.