FridgeIQ (Kunal & Rudy)

Inspiration

What it does

How we built it

Challenges we ran into

Here's your full Devpost "About the project" writeup (it follows Devpost's standard sections) plus the Built With list — all paste-ready:

Inspiration

Every family knows the moment: it's 6pm, the fridge is full, and nobody knows what's for dinner. The USDA estimates 30–40% of the US food supply is never eaten — roughly $1,500 a year for an average family — and most of it was perfectly good the day someone forgot it existed. Recipe apps exist, but they all start with typing. Nobody types their fridge. We wanted dinner to start with a photo.

What it does

FridgeIQ is an AI food brain for your whole household. Snap a photo of your fridge, a handwritten leftovers note, or a grocery-store flyer — AI vision turns it into an editable ingredient list, flagging what should be used soon and what it's unsure about. One tap generates three recipes ranked to rescue expiring food first, each with a Waste Saver Score, cost per serving, macro estimates, substitutions, and a tap-to-check-off cooking mode.

What makes it different: Flyer intelligence reads "chicken $2.99/lb" off a flyer photo and merges the deal into your kitchen, pricing dinner with it. Picky Eater Mode adapts every recipe when you say "my 6-year-old won't touch anything green" — sneak-the-veggies-in tips included. Nutrition goals treat "40g+ protein, under 600 kcal" as a hard constraint, adjusting portions until the numbers comply. And the signature Leftover Loop: every dinner predicts its own leftovers and plans tomorrow's lunch from them. One scan, two meals. There's also a smart shopping list (missing ingredients + sale prices), saved recipes, and an impact dashboard backed by a shared Postgres ledger showing all-time community stats.

How we built it

Next.js 16 (App Router) with strict TypeScript and Tailwind CSS v4. All AI calls run server-side through API route handlers — the browser never sees a key. Claude's multimodal API handles vision (fridge photos, handwriting OCR, flyer extraction) and recipe generation, both under strict JSON contracts. A dedicated validation layer parses, clamps, and sanitizes every model response before it touches the UI — malformed output degrades to a friendly fallback, never a crash. Images are client-side downscaled before upload to keep vision calls fast and cheap. A Supabase Postgres ledger (row-level security, server-only keys, fire-and-forget writes) powers community impact stats. Deployed on Vercel.

Challenges we ran into

Real-device testing surfaced real bugs: crypto.randomUUID doesn't exist on plain-HTTP local networks (crashed the app on phones — fixed with a safe fallback), cached recipes went stale when the kitchen or preferences changed (fixed with content fingerprinting), and AI recipe steps were being truncated mid-word by a validator meant for ingredient names. Getting the AI to treat macro goals as hard requirements — not suggestions — took prompt engineering with explicit numeric constraints and honest "closest possible" annotations when ingredients can't reach a target.

Accomplishments that we're proud of

The demo cannot die: a three-tier fallback (real AI → deterministic sample data → preloaded demo fridge) means the full experience survives lost wifi, missing API keys, or provider outages. We're also proud of the honesty-first design — low-confidence detections are flagged "AI unsure, please check," and every score, cost, and macro is labeled an estimate.

What we learned

Never trust raw model output — validate everything at the boundary. Design for failure before designing features. And the best features came from bugs found while pointing the app at a real fridge, real handwriting, and a real phone.

What's next for FridgeIQ

Expiration tracking with push notifications ("use your spinach today"), a nutrition API for verified macros, receipt scanning, household profiles with cross-device sync, weekly meal planning, and grocery integrations — Instacart/Walmart checkout for a recipe's missing ingredients is one integration away, since we already connect store deals to dinner.