About the project

CookTheFeed turns short-form cooking videos into structured recipes you can actually cook.

Inspiration

I was inspired by Eitan Bernath's energy and by a clear gap in modern food content: people discover amazing recipes on TikTok, YouTube Shorts, and Instagram Reels, but rarely cook them because the instructions are buried in fast, chaotic videos. I wanted to close that gap by turning video inspiration into clear, practical cooking steps with AI.

What it does

• Accepts a video URL (or Android Share Intent) from TikTok, YouTube Shorts, and Instagram Reels.
• Core UX is optimized for these three platforms; additional providers supported by yt-dlp (for example, Facebook video links) are available on a best-effort basis.
• Uses AI to extract dish name, ingredients, instructions, prep time, difficulty, servings, calories, cuisine, dietary tags, and spice level.
• Detects allergens and suggests smart substitutions (Pro).
• Lets users save recipes into a personal cookbook with search, filters, and shopping checklist mode.
• Keeps attribution by linking users back to the original creator video.
• Includes a trending feed so users can experience value immediately.

How I built it

I built CookTheFeed as a solo project end-to-end:

• Mobile app: React Native + Expo (Android-first in this phase), Expo Router, NativeWind.
• Backend API: FastAPI (Python) with extraction, caching, and credit/refund logic.
• AI extraction: Google Gemini 2.0 Flash for multimodal video analysis.
• Data/auth/storage: Supabase (Postgres, Auth, Storage).
• Monetization: RevenueCat subscriptions.
• Reliability and observability: yt-dlp with Cobalt fallback, Sentry monitoring.

Challenges we faced

As a solo developer, the biggest challenge was balancing product quality, reliability, and speed at the same time.

• Social video extraction reliability changes often, so I implemented a dual download strategy (yt-dlp primary, Cobalt fallback).
• I implemented thundering-herd protection so the same video is not extracted multiple times in parallel during concurrent requests.
• Abuse prevention required multiple fraud-protection layers so users could not game AI extractions through retries, account switching, or edge-case request patterns.
• Building fair usage controls required strict credit accounting, refund paths, and rate limiting.
• Anonymous-to-Google account merge and subscription restore flows required careful handling for real mobile behavior.
• AI latency required polished loading UX, cancellation support, and cache-first retrieval.

What we learned

• Reliability and trust are as important as AI output quality.
• In AI products, model accuracy is only half the job; robust guardrails and anti-abuse controls are equally important.
• A focused Android-first rollout helps validate core behavior faster.
• Building production-minded architecture as a solo founder requires aggressive prioritization and clear technical boundaries.

Current status

The app is currently in Android closed testing and is fully sufficient for competition participation, with a clear roadmap after the event.

What's next for CookTheFeed

• Multilingual extraction and localization, so users can process videos in one language and receive recipes in their preferred language.
• Smarter personalization with stronger dietary intelligence and preference-aware recommendations.
• Expanded Pro capabilities focused on safety, including richer allergen insights and more actionable substitutions.
• iOS release after Android closed testing, with a shared backend and polished cross-platform experience.
• Continued platform hardening with stronger abuse resistance, extraction integrity checks, and reliability improvements as usage scales.

Built With

• cobalt
• expo.io
• fastapi
• googlegemini-2.0-flash
• nativewind
• python
• railway
• react-native
• revenuecat
• sentry
• supabase
• yt-dlp