Shelfy: Tomorrow's Best Sellers
AI Demand Prediction - Rank products by tomorrow's conversion rate. Top 30% products captures 50% of actual sales.
Usage
- Enter today's product stats
- See tomorrow's predicted conversion + revenue
- Stock Top 10% products → 2.5x ROI
Business Results (42M events test)
- Top 10% → 25.4% sales - ROI: 2.5x
- Top 30% → 52.1% sales - ROI: 1.7x
- Test: Oct 25-29 (blind) Dataset: 42M ecommerce events
Features
- Live prediction - Tomorrow's conversion rate per product
- Demand signals - High / Medium / Low
- production features - Views, carts, momentum, intent ratios
Model Card
- Architecture: XGBoost Regressor
- Features (8): views(29%) + carts(29%) + cart momentum(13%) + user metrics(17%) + other(2%)
- Target: Tomorrow's purchase rate (actual purchases / views)
- Stability: 82% day-to-day Top 10% products overlap
Production Stack
- Data: Clickstream (views, carts, purchases)
- Stack: Python + XGBoost + FastAPI + PostgreSQL
- Retraining: Weekly with shelfy's real data
Tech Stack
- Model: shelfy_purchase_rate_v2.joblib (XGBoost)
- UI: Gradio 4.44.0
- Data: Ecommerce behavior (Oct 2019)
Built for datastorm hackathon | Apache 2.0 | Deployed 1/16/2026
