FridgeChef

FridgeChef answers the daily 'what's for dinner?' question with the least possible input: open the app on your phone at 17:30, photograph the open fridge, and get three concrete dinner ideas ranked by how completely they can be cooked with what is actually there. Each suggestion carries a short step-by-step recipe, a realistic cook time, and a missing-items line ('cream would improve the sauce').

Recognition and generation are deliberately split into two AI calls. A vision model reads the photos once and returns the ingredients it found as editable chips, so recognition mistakes are corrected by a human before any recipe is written, and the expensive vision call never runs twice. A text model then writes exactly three recipes, honoring the household's preferences: diet, allergies as hard constraints, dislikes, and household size, all stored locally in the browser.

The UI is Danish-only by design (it is a household tool first) and leans on Danish everyday cooking, down to knowing what leverpostej is. It ships as an installable PWA with a mobile camera flow, client-side image compression before upload, no photo retention on the server, and cost guards in the form of size caps, rate limiting, and a spend-capped API key.