✦

The Problem: Data Graveyard

I kept saving recipe videos on YouTube, but the playlist quickly became a graveyard — a growing collection where inspiration was forgotten. I found myself scrolling through a flat list of thumbnails, unable to tell if a recipe fit my needs, and usually gave up to search for something new instead.

✦

The Concept

I built RecipeBrain to turn a list of bookmarks into an intelligent kitchen assistant that understands the context behind every video, the project focuses on three interactions:

Agentic Discovery: Chat with an AI assistant to filter, compare, and recommend saved recipes by ingredients, time, or mood.

Automated Extraction: importing a recipe from a YouTube link with AI-extracted metadata.

Smarter Filing: browsing and filtering a structured collection by time, difficulty, cuisine, and dietary needs.

✦

How It Was Built — With AI

Step 1: Design in Figma

Natural hand gestures control the AR experience — diners performs an opening hand gesture to activate the menu, close it to dismiss, and use grabbing motions to select items.

Step 2: Translate Design to Code

Use Claude Code + FigmaMCP to convert Figma designs into React components, keeping implementation aligned with design tokens.

Step 3: Define the Technical Architecture

Use Claude to work through stack decisions, data structure, AI prompt architecture, and implementation sequence to produce a technical brief that guides the build.

Step 4: Wire Up Data and AI Chat

Connect pre-seeded recipe JSON to UI components. Implement the AI chat using Gemini 2.5 Flash, passing the full collection as context with each prompt.

Step 5: Deploy to Vercel

Push to GitHub and deploy via Vercel with the Gemini API key as an environment variable.

Step 6: Add YouTube Import Feature

Build a serverless function that extracts a YouTube transcript, parses it into structured recipe data via Gemini, and wires it into the existing Add Recipe flow. Capped at 2 imports per session.

Step 7: Evaluate + Refine with Braintrust

Use Braintrust (an AI evaluation platform) to test chat accuracy and import parsing quality. Refine system prompts based on results.