A fault-tolerant streaming anomaly detection platform built for real-time fraud monitoring — with Kafka, MLflow, drift detection, and adaptive retraining pipelines.
RealSignal is a streaming-first anomaly detection system focused on operational reliability, lifecycle monitoring, and adaptive ML infrastructure — not just model training.
Kafka-driven real-time transaction processing with online feature engineering and low-latency anomaly detection.
Dead Letter Queue handling isolates malformed events and prevents consumer crashes during inference failures.
PSI-based drift detection continuously monitors changing transaction distributions and model degradation.
Complete ML lifecycle orchestration with full operational observability at every stage of the pipeline.
Velocity tracking, merchant diversity analysis, and rolling customer transaction statistics.
Rule-based anomaly explanations provide interpretable reasoning behind fraud predictions.
Experiment tracking, metric logging, drift reports, and retraining lifecycle management.
Significant feature drift automatically triggers retraining on updated transaction distributions.
FastAPI endpoints expose health checks, drift reports, and lifecycle observability.
Drift monitoring and evaluation artifacts help identify model degradation over time.
Streaming-first architecture focused on reliability, monitoring, and adaptive retraining workflows.
Simulate transaction behavior and observe real-time anomaly detection with explainability signals.
Amount: 1500
Velocity: 20
Avg Amount: 4500
Merchant Diversity: 3
Amount: 90000
Velocity: 20
Avg Amount: 40000
Merchant Diversity: 4
Isolation Forest identifies anomalies based on statistical rarity and density patterns learned from training data.
Transactions that fall outside learned feature distributions may be flagged as anomalous even if they appear normal to humans.
Run a transaction analysis to view results.