RésuméEmail me

ForkLog Live

Turn a viral cooking video into a recipe, then cook it with a camera and a voice coach that notices what you did differently.

Year
2026
Role
App, prompts, edge functions, AI tuning
Team
Hackathon pair
Stack
Expo, Gemini, ElevenLabs, Supabase edge functions
Link
Repository on GitHub

How it works

Filed as version 2

v2 keeps about 3 tbsp soy and 90 s garlic

Garlic soy butter noodlesCreatorYou
Soy sauce, step 42 tbspabout 3 tbsp
Garlic, step 360 s90 s

Rating 8 of 10Version v2

SpokenVisualInferredSpoken, visual or inferred: the recipe tells you which.

The system

From a cooking video to version 2 The Expo app uploads a cooking video to Supabase Storage and calls the extract-recipe edge function, which sends the whole video and audio to Gemini Pro and saves version 1 with every quantity tagged spoken, visual, text or inferred; while cooking, changed camera frames go to coach-frame, which compares them with the creator's reference frame on Gemini Flash, and the correction is spoken through the tts function and ElevenLabs; the recap function returns a diff against the creator and the app saves it as a new version, while model keys never leave the edge functions. Model keys live only here, as Supabase secrets loop each changed frame, at least 2 s apart upload video extract-recipe video + audio each quantity tagged spoken, visual, text or inferred recipe, version 1 coach-frame frame + reference coach event tts, correction speak line audio, played recap + rating diff vs creator save as new version Expo app camera and speaker Supabase data Postgres, Storage Edge functions Deno, hold the keys Gemini Pro video, Flash frames ElevenLabs Flash v2.5 voice Call Return Cook saved as a new version Key boundary

The idea

Cooking videos are fast and vague. ForkLog reads the whole video, sound included, and writes down the recipe with quantities, tagging each one by how it was found: spoken, visual, on screen text, or inferred.

ForkLog library with a saved recipe and its versions
ForkLog recipe extracted from a video, quantities tagged spoken, visual, text or inferred

Cooking with a coach

During the cook the camera checks each step and a voice coach answers. When you add more soy sauce than the creator did, it notices and suggests a splash of water.

ForkLog cook screen, step one
ForkLog coach suggesting a splash of water after extra soy sauce

The recap

Afterwards you get a diff of your cook against the creator's, and can save it as a new version of the recipe. Model keys live in edge functions, never in the app.

ForkLog recap: your cook against the creator
ForkLog version history with v1 and v2
Next, uab bank apps →