An enterprise-grade, event-driven anti-money laundering (AML) platform designed to detect synthetic transaction injection, localized velocity spikes, and organized mule accounts in real time.
- Row-Level ML (Isolation Forest): Evaluates incoming transaction volumes and rolling transaction frequency to score statistical anomalies.
-
In-Memory Velocity Tracking (Redis): Tracks sliding-window account velocity (
$t-60\text{s}$ ) using RedisZSETaggregations for sub-millisecond lookups. - Graph Topology & Mule Hub Detection (NetworkX): Evaluates directed money flows to identify accounts with high in/out degree parity operating as high-throughput pass-through conduits.
- Macro Community Detection (Louvain Method): Dynamically partitions the global transaction graph into modularity-optimized clusters to isolate coordinated fraud rings.
-
Macro Audit (Benford's Law): Evaluates empirical leading-digit distributions against
$P(d) = \log_{10}(1 + 1/d)$ via Kolmogorov-Smirnov statistical divergence tests to detect automated wash-trading scripts.
- Event Ingestion: Redpanda (Kafka-compatible event broker)
- Real-Time State: Redis (In-memory sorted sets for sliding window counts)
- Storage & Audit: PostgreSQL (Immutable compliance ledger)
- Analysis & ML: Python, Scikit-Learn, NetworkX, python-louvain, Scipy
- Dashboards: Streamlit (SOC UI with interactive PyVis network graphs) & Metabase (Operational BI)
- Boot the complete infrastructure:
docker-compose up -d --build
- Launch the synthetic fraud & mule generator:
docker exec -it fraud_api python src/producer.py - Access the Interfaces:
- Streamlit SOC Dashboard:
http://localhost:7501(Interactive PyVis graph, Benford curve, live alerts) - FastAPI Docs & Metrics:
http://localhost:8000/docs - Metabase BI:
http://localhost:3000
- Streamlit SOC Dashboard: