try it

stopeating.pizza

drop in a photo of a grocery receipt (or a screenshot of an online order) and it grades your cart. no account needed to poke at the sample. yes, the name is a little judgmental. that's the point. the domain isn't live yet, so for now the screenshots below are from a local build.

overview

stopeating.pizza reads a photo of your grocery receipt, or a screenshot of an online grocery order, and hands you back a health report. It pulls out every line item, matches each one to real nutrition data, and scores it. You get an A to E grade on every item (the same system the food labels use across Europe), one overall score for the whole cart, the actual numbers behind every flag, one better thing to reach for next time, and a flag on anything in your cart that Europe restricts or bans. The whole thing is built around improvement instead of shame. I wanted it to read like a friend who happens to know nutrition looking over your cart, not a calorie counter yelling at you.

A stopeating.pizza report: a 7.9 out of ten cart score, 16 items read, 8 worth changing and 8 in the clear

the report

a grade on every item: each item gets an A to E letter grade based on the official French Nutri-Score. The card is bordered in its grade color, green through red, and you tap it open to see the real per-serving numbers, the source behind each flag, and one specific swap worth making.
a cart score: one number out of ten for the whole cart, weighted by how much of each food you actually buy, not by how many items landed in the basket. A bag of spinach and a bag of candy do not count the same.
it shows its work: every flag points at the body that sets the limit (the WHO, the FDA, and friends), and it never guesses. If it cannot confidently identify a product, it leaves it out of the score and tells you, rather than making a number up.
banned there, sold here: some of what's legal on American shelves is restricted, warning-labelled or banned outright in Europe. The report flags anything in your cart on that list, and the full list has its own page: every entry names the regulation, says when a restriction is precautionary rather than a finding of harm, and calls out popular claims that are wrong.
recall guardian: every upload is checked against active FDA and USDA food recalls, so if something in your cart got pulled, it shows up as a warning on the report.
your next cart: the report doubles as next week's shopping list. Keep the winners, swap the rest, copy the list.
reproducible forever: every report is stamped with the exact engine and rubric versions that produced it, so the same receipt reproduces the same report even after the scoring rules change.
An opened item card: an E grade, the per-serving numbers, each flag with its FDA source, and a better thing to reach for The banned list: 26 things in American food that Europe restricts, each naming the regulation

construction

TypeScript Next.js pnpm monorepo Supabase PostgreSQL Anthropic API Tesseract OCR USDA FDC Open Food Facts Vitest Vercel

The scoring engine is a standalone TypeScript library. It knows nothing about the web, the database, or who you are: food items go in, scored records come out, and everything it needs (the nutrition databases, the OCR, the vision model) is handed to it as an injected interface. A hard rule keeps cached facts (nutrients, ingredients) separate from recomputed judgments (grades, scores, flags), which is what lets every report stay versioned and reproducible.

Reading a receipt is cost-aware. Free on-device OCR (Tesseract) runs first, and it only pays for a vision model when the OCR is not confident enough. Nutrition facts come from USDA FoodData Central and Open Food Facts, and each product is resolved once and cached, so the second time anyone buys it the lookup is free.

The grade uses the official French Nutri-Score algorithm (the 2023 version). I did not want to just claim it was right, so I wrote a test that runs my grader against Open Food Facts' own published grades: it matches on 100% of 728 real products, and it stays in the suite as a permanent regression test. Building it caught several real bugs in my scoring tables, which was exactly the point.

architecture

receipt photo → OCR / vision read → resolver (USDA + OFF, cached) → scoring engine → Next.js report → Supabase (Postgres + RLS)

The web app is Next.js on Vercel. Supabase handles auth, Postgres, and file storage, with row-level security so one person can never open another person's report. The engine lives as its own package in a pnpm monorepo, covered by around 120 tests written test-first, so a change to the scoring rules cannot quietly break a grade.

Source? It lives in a private repo, since this one is a collaboration. The rest of my work is on my GitHub.