A personalized recipe recommendation web app. It learns your taste from ratings, knows what's in your fridge, and respects your dietary restrictions.

• Personalized recommendations — the more recipes you rate, the better it gets
• Fridge-aware — tell it what ingredients you have and it prioritizes recipes you can actually make
• Dietary filters — filter by vegan, gluten-free, and other diet labels
• Search — keyword search and semantic (meaning-based) search powered by sentence embeddings
• Cold-start support — works even for brand new users by falling back to popular recipes

Recommendations go through three stages:

1. Retrieval — a Two-Tower neural network scores all recipes against the user's taste profile and returns the top 100 candidates. New users fall back to a popularity ranking based on Bayesian average rating.

2. Ranking — a LightGBM model re-scores the candidates using features like nutritional fit, fridge overlap, and retrieval score.

3. Diversification — Maximal Marginal Relevance (MMR) picks the final results, balancing relevance with variety so you don't get ten pasta dishes in a row.

For a full walkthrough of the data pipeline, model training, and evaluation see notebooks/notebook-picky-eater.ipynb.

1. Install dependencies

uv sync

2. Add your artifacts

The app expects trained model files in an artifacts/ folder at the project root. This folder is not included in the repo (it's too large). Train the models using the notebooks in notebooks/ or get the artifacts separately.

3. Start the server

uv run uvicorn app.main:app --reload

Then open http://localhost:8000 in your browser.

Made for food lovers, picky eaters, and everyone in between :)