feature-engineer

Streaming transaction features are computed from Kafka events by using Faust for asynchronous ingestion and publishing enriched records. The service parses JSON events, calculates per-user 60-second statistics, and publishes enriched fraud-detection features downstream.

A Redis sliding window calculates per-user transaction statistics by maintaining a 60-second amount window for each user. This pattern enables the real-time computation of mean, standard deviation, and z-scores to detect anomalous transaction behavior.

Risk features are derived from raw transaction streams by evaluating merchant risk, odd-hour activity, and composite risk scoring for each event. This transforms raw Kafka transactions into model-ready inputs for real-time anomaly detection.

Robust Kafka I/O handling and error management are achieved through native health and metrics endpoints. The service integrates Prometheus metrics to monitor reliability and handle ingestion errors within the Faust stream processing pipeline.

Faust and Redis can be used together for real-time anomaly detection pipelines. Faust handles asynchronous Kafka event consumption while Redis provides the backend for sliding window feature computation required for reliable streaming fraud detection.

Accurate JSON event parsing is required in a feature engineering pipeline because raw Kafka transactions must be precisely decoded to compute per-user statistics and risk features. Parsing failures disrupt the generation of model-ready inputs for anomaly detection.