Splitbill.

Splitbill takes the most tedious part of eating out with friends — working out who owes what — and turns it into a thirty-second job. You snap a photo of the receipt, AI reads every line item and price, and you assign each item to people with a tap or by typing the way you'd actually say it: "paneer for Aman, water for everyone". It does the tax and tip math, then hands you one read-only link to share with the table.

It's a PWA, so it installs like a native app on iOS and Android, works without a signup to start, and lives at bill.rajkar.dev.

Everything — product, design, frontend, backend, the AI pipeline, the credit system, and deployment. It's a solo project, which is the point: it's where I own every decision end to end, from the receipt-scanning prompt to the share-link data model.

Reading a crumpled restaurant receipt reliably. Photos are skewed, blurry, in mixed currencies, and the line items follow no standard. The scan runs through Gemini with a strict output schema, so the model returns structured items — name, quantity, price — instead of prose I'd have to parse. Natural-language assignment ("paneer for Aman") goes through Groq's Whisper for transcription and then a second AI pass that maps spoken names onto the actual people and items on the bill.

The other half is making something AI-powered cheap enough to give away. Every scan costs money, so there's a credit system — free credits on signup, top-up packs after — with failed AI runs auto-refunded, so a bad photo never costs the user a credit.

Next.js 16 with the App Router and React 19, Supabase for auth and data, Tailwind for styling. Each bill is stored once and exposed through two links — an editable owner link and a read-only share link — so whoever scanned keeps control while everyone else just sees the split.

Open Graph images are generated per bill, so a shared link previews the real total and who owes what instead of a generic card. The whole thing ships as a PWA with offline-aware caching, deployed on Vercel.

Splitbill is live at bill.rajkar.dev and does exactly what it set out to: a real receipt becomes a shareable, itemized split in under a minute, with nothing to install and no signup to start. It's also been my testbed for production AI-product patterns — structured-output prompting, credit metering, per-bill OG generation — the kind of work that increasingly shows up in client projects too.