Reles
Saved to Served
Cooked For You
One Tap Grocery Lists

Reles

Inspiration

We kept noticing the same pattern: find an incredible recipe on YouTube, save it, never cook it. The friction between watching and doing—pausing with messy hands, estimating measurements, scribbling ingredients—kills momentum. The problem wasn't motivation, it was format. Reles was born from a simple question: what if saving a recipe actually meant you could cook it?

What it does

Reles transforms YouTube cooking videos into structured, actionable recipes in seconds. Paste a link and the app extracts ingredients, instructions, cook time, allergens, difficulty, and nutrition—streamed live to the screen. Preferences shape a personalized discovery feed. Saved recipes convert into aisle-organized grocery lists with ingredient aggregation across meals and one-tap Instacart ordering.

How we built it

Mobile: React Native + Expo with Expo Router, RevenueCat for subscriptions
Backend: Node.js/Express on Railway with Firebase Firestore

The extraction pipeline fetches YouTube captions via SerpAPI, passes them to Gemini 2.5 Flash, and streams structured JSON back to the client through Server-Sent Events. The UI renders ingredients and steps progressively as they arrive rather than waiting behind a loading spinner.

Challenges we ran into

Streaming JSON parsing — LLMs don't output valid JSON token by token. We built an incremental parser that detects completed objects mid-stream and emits them as typed SSE events.

Quantity normalization — Merging "a handful of basil" and "½ cup flour" across recipes into a coherent grocery list meant parsing fractions, unit aliases, and ambiguous quantities into a common format.

Transcript reliability — Not every video has usable captions. We had to filter search results and build fallback handling to balance coverage against extraction quality.

What we learned

Progressive UI feedback matters more than raw speed—users perceive a 10-second extraction with animated reveals as faster than a 7-second one behind a spinner. Prompt engineering for structured extraction is its own discipline; our extraction prompt went through dozens of revisions to consistently produce valid JSON. And building a pluggable LLM provider interface early paid off when we needed to switch providers mid-development.

What's next for Reles

Multi-video recipe synthesis, a cooking mode with timers and voice guidance, ingredient substitutions based on pantry, and platform expansion to Android, browser extensions, and TikTok/Instagram support.