What it is
Production-style e-commerce recommender stack that covers data pipelines, feature stores, BST training, Triton serving, rollout gates, and observability on Kubernetes. Builds and exercises a full recommendation platform from event generation to personalized Top-K serving, with batch/stream features, BST training, governed deployment, and online rollout control.
Key results
- Streams PostgreSQL CDC via Debezium/Kafka into Flink and Redis/Feast
- Uses BST training with Kubeflow, Ray Tune, MLflow, ONNX, and Triton
- Rolls out candidates at 10%→25%→50% traffic with auto rollback
Shows that a recommendation system can be built as a fully governed, real-time MLOps platform rather than a standalone model.