E-commerce Recommendation Pipeline

Proof of concept: a live feature store backed by a public product catalog. The service loads the feature table into memory at startup and serves product-level recommendation signals through a tiny API.

Model input Product features

Recency, price, review, catalog, and categorical signals

Storage Parquet

Feature table is precomputed and served from memory

API FastAPI

GET /health and GET /features/{entity_id}

How a request flows

Step	What happens
1. Ingest	Fetch products from the public API and store timestamped raw JSONL.
2. Clean	Normalize text, timestamps, and numeric fields; deduplicate by product ID.
3. Engineer	Create recency, frequency, categorical, and scaled numeric features.
4. Serve	Load the parquet file at startup and return a product feature vector instantly.

Example output

{"entity_id": 1, "features": {"price": 9.99, "rating": 4.6, "days_since_updated": 2.1, "review_count_scaled": 0.84}}

Endpoints: /health and /features/{entity_id}