Personalized Recommendations & Demand Forecasting

Problem

E-commerce stores struggle with two related problems: surfacing the right product to the right customer (recommendations), and stocking the right amount of inventory ahead of demand (forecasting). Most off-the-shelf systems treat these in isolation: recommenders cold-start poorly on new customers, and forecasts ignore the personalization signal entirely.

Approach

A hybrid pipeline that fuses content-based and collaborative filtering for recommendations, then routes the aggregate demand signal into a per-SKU forecasting model. SQLite stores the rolling order history; pandas handles feature engineering; scikit-learn trains the underlying models. A Streamlit interface lets a non-technical operator browse predictions per user, per product, and per category.

Result

A working end-to-end system that demonstrates how a single model surface can serve both merchandising and operations, and how addressing cold-start at the recommender layer feeds cleaner signals back into the forecast.