akiragy/recsys_pipeline

# Technical Summary Implements a complete recommender system pipeline using two-stage retrieval: term-based and vector-based recall via Elasticsearch, followed by DeepFM ranking with 59 engineered features (one-hot, multi-hot, and dense). Enforces point-in-time joins during feature generation to prevent leakage while maintaining offline-online consistency. Containerizes all production components (Redis for user vectors, Elasticsearch for ANN item retrieval, Feast for feature serving, Triton for inference) in Docker while training occurs in Conda, demonstrating an end-to-end workflow from data preprocessing through online serving on a single machine.

No commits in the last 6 months.