MissingBite

Your handwritten recipe, understood by your confirmed kitchen.

Comment

Inspiration

Most recipe tools start with a generic catalogue. MissingBite starts with the food and recipes a household already has. The idea grew from trying to inspect a fridge during a live visual conversation: that test showed that guided still images, visible uncertainty, and user confirmation are more trustworthy than pretending a continuous stream is understood perfectly.

What it does

MissingBite is an AI-native personal kitchen platform:

  • Photograph one to five useful views of a fridge, freezer, pantry/cupboard, or another kitchen area.
  • Review, correct, confirm, or reject GPT-5.6 item candidates before anything becomes current inventory.
  • Import one handwritten recipe from one to eight ordered pages, preserve every Original page, and create an editable normalized Cook Mode.
  • Create a structured recipe manually by typing or voice when there is no scan.
  • Deterministically classify every recipe ingredient as Have, Check, or Missing.
  • Add all missing ingredients to an editable shopping list in one reviewed batch.
  • Review checked purchases, choose where each belongs, and transfer only confirmed purchases into current inventory; recipe availability then recalculates.
  • Use a global live voice companion for hands-free Cook Mode, navigation, and app actions that still pause for visible approval.
  • Enter a clearly labelled, resettable Seeded demo when a user wants to explore without an account or a live model dependency is unavailable.

The live web release candidate is available at https://missingbite-web.onrender.com.

Trust is part of the product

AI results enter candidate or draft records, never confirmed household data directly. Unknown or unreadable values remain unknown. Recipe originals stay immutable while Cook Mode remains editable. Matching is conservative, and a shopping checkmark never silently becomes inventory.

Authenticated records are protected by Supabase row-level security, and recipe images remain in private Storage. MissingBite does not claim freshness, expiry, allergen safety, medical suitability, or exact remaining quantity from an image.

How it was built

The client uses Expo, React Native, and TypeScript. A FastAPI/Python service protects the live model routes. Supabase provides Postgres, Auth, private Storage, and row-level ownership policies. Render hosts the web client and API.

GPT-5.6 Terra performs live multimodal kitchen-photo analysis and handwritten-recipe transcription through the OpenAI Responses API. Requests use original image detail, strict versioned structured outputs, high reasoning for kitchen scans, and xhigh reasoning for harder handwritten recipes. Deterministic application code—not the model—handles matching, arithmetic, state transitions, and shopping-list generation. Hands-free conversation uses the OpenAI Realtime API.

Codex acted as the engineering collaborator across product scoping, architecture, Expo/FastAPI implementation, strict schemas, Supabase migrations and RLS, tests, CI, deployment, rendered QA, security review, and evidence documentation. Human decisions defined the product boundary and kept uncertain food data under user control.

Build Week contribution boundary

The kitchen-platform idea, household requirements, and an initial planning handoff existed before Build Week. The implementation is the Build Week extension: the repository history begins July 16, 2026 and contains the Expo client, FastAPI service, GPT-5.6 contracts, Supabase data and authorization layer, persistence, tests, CI, deployment configuration, and release evidence built with Codex during the event window.

Try it

Open https://missingbite-web.onrender.com in a clean browser and choose Seeded demo for the no-account path. Follow Kitchen → Review → Cookbook → Cook → Shopping → reviewed transfer back to Kitchen.

The current competition build deliberately uses bounded still images rather than continuous video. Nutrition optimization, product-label comparison, retailer checkout, arbitrary recipe-website import, and ChatGPT App/MCP access remain post-competition roadmap items.

Built With

Share this project:

Updates