Forklog Live

Inspiration

We save cooking videos on TikTok and Instagram, then fail to cook them. The recipe is buried in fast cuts, quantities are never said out loud, and nobody watches your pan while you cook. We wanted the creator to feel like they were standing next to you.

What it does

ForkLog Live turns a viral cooking clip into a recipe you can cook along to.

Import a cooking video. Gemini watches and listens, then writes the ingredients and steps.
Cook with your phone camera on. The app compares your pan to the video, speaks each step, and corrects you out loud ("that's too much soy").
Recap after the meal. The app shows what you changed and saves it as a new recipe version, so v2 is your version.

How we built it

App: React Native with Expo and TypeScript.
Backend: Supabase for the database, video storage, and Edge Functions.
AI: Gemini reads the video and the live camera frames. ElevenLabs gives the coach its voice.
Safety: No API keys live in the app. Every AI call goes through an Edge Function, and every model answer is checked with zod before the app uses it.

Challenges

Latency and cost: Sending camera frames constantly is slow and expensive. We only send a small frame when the scene changes, at most request at a time.
Unreliable models: Gemini can burn broken JSON. We added a fallbackchain across models, retries, and strict validation.
A talkative coach is an annoying coach: We built an alert gate so the same warning is never spoken twice within 20 seconds.
Food safety: The coach never says raw meat, poultry, or eggs are "done" from vision alone.