At Andersen, I developed ETL components for the A-Benefits personalized discount recommendation system, moving user interactions from PostgreSQL into recommendation data and loading Top-N results back into the database.
I implemented a cold-start fallback for new users and a separate recommendation path for users with interaction history. I also built LightGBM ranking features and worked with a hybrid stack combining content-based filtering, SVD/ALS, and LightGBM re-ranking.
I integrated AWS S3 for ML artifacts and feature files, prepared a FastAPI online-inference skeleton, and benchmarked Qwen3 and Mistral via Ollama for structured JSON quality and inference latency.
