Title: Fridge-to-Recipe (leftovers feature) Problem: A user standing in front of their fridge at 6pm wants to cook tonight with what they already have, but the app currently only helps when they know what to search for. They don't want to type in every ingredient. Goals: - A user can go from photo to a usable recipe list in under 30 seconds. - At least 60% of suggested recipes use only detected ingredients plus common pantry staples (salt, oil, etc.). - The feature works without any manual ingredient typing in the common case. Non-goals: - Inventory tracking over time (what's in the fridge across days). - Grocery list generation or shopping integration. - Nutrition or calorie analysis of the suggestions. - Barcode scanning of packaged goods. Users and use cases: - Weeknight cook at 6pm deciding dinner from what's on hand. - User avoiding a grocery trip who wants to use up perishables before they spoil. Requirements: - MUST: capture or upload a fridge photo and detect distinct food items from it. - MUST: match detected items against the recipe database and return recipes rankable by "fewest missing ingredients." - MUST: let the user remove or add a detected item by tapping, for when detection is wrong. - SHOULD: assume a default pantry (salt, pepper, oil, common spices) so recipes aren't blocked by staples. - COULD: let the user save a detection result to revisit later in the session. Acceptance criteria: - Given a photo containing recognizable items, the app returns at least one recipe using only detected items plus default pantry, or clearly states none were found. - A user can correct a misdetected item and the recipe list updates without re-taking the photo. - The flow from photo capture to first recipe list completes without any text entry. Assumptions: - The recipe database is tagged with structured ingredient lists the matcher can query. - An image-recognition model (built or third-party) can identify common fridge items with usable accuracy. - Users will accept approximate detection if correcting it is one tap. Open questions: - What detection accuracy is good enough to ship, and who decides the threshold? (Needs a labeled test set.) - Build, buy, or call an API for food recognition? (Cost and latency tradeoff, needs eng input.) - How are quantities handled, or is presence/absence enough for v1? Risks: - Detection accuracy is the whole feature; if it's poor, the experience fails. Early signal: detection precision on the test set below the chosen threshold. - Latency on image processing pushes past the 30-second goal. Early signal: median photo-to-results time in testing.