FeatureForge

ML Feature Store & Drift Detection

An enterprise feature store on Snowflake with point-in-time-correct retrieval, feature versioning, and lineage — paired with an automated drift-detection layer wired into SageMaker Model Monitor.

Methods & contributions

• Point-in-time correctness — training-serving skew avoided through as-of joins and versioned feature definitions.
• Drift detection — textbook Population Stability Index (PSI), Kolmogorov–Smirnov, and chi-squared tests with standard severity bands (PSI 0.1 / 0.2 / 0.3).
• Orchestration — Airflow DAGs schedule feature materialization and drift scans.
• Model monitoring — SageMaker Model Monitor integration with automated retraining triggers.

Data flow

sources → Snowflake feature tables (versioned)
        → point-in-time retrieval → training / serving
        → drift detector (PSI · KS · chi²) → SageMaker Model Monitor

Technologies