Brian-10-star/mpesa-fraud-detection

# M-Pesa Fraud Detection ML Platform ## Project Highlights - Built an end-to-end M-Pesa fraud detection platform using Kafka, PostgreSQL, FastAPI, MLflow, and Docker - Simulated realistic transaction streams across 58 Kenyan cities - Engineered 20 fraud features spanning behavioral, temporal, velocity, and amount patterns - Added model monitoring, prediction logging, and drift detection - Automated testing and CI validation using GitHub Actions An end-to-end ML platform for near real-time fraud detection on M-Pesa transactions. This project demonstrates production-style data engineering across streaming ingestion, feature engineering, model training, API serving, drift monitoring, and automated testing, built with a Kenyan fintech context using realistic M-Pesa transaction types, Kenyan names, KES amounts, and location data across 58 Kenyan cities. --- ## Platform Preview **Streamlit Dashboard** **MLflow Experiment Tracking** **FastAPI Interactive Docs** --- ## Business Problem M-Pesa processes millions of financial transactions daily. Fraud patterns, including velocity attacks, account takeovers, and suspicious high-value transfers, evolve continuously and require automated detection systems that can score transactions in near real time. This platform detects: - Unusual spending behaviour relative to a sender's historical baseline - Rapid transaction bursts within short time windows - Account takeover signals via new device and new location combinations - Statistically extreme amounts using z-score analysis --- ## Architecture The platform is built in 8 layers, each feeding the next: --- ## Tech Stack | Layer | Technology | |---|---| | Streaming | Apache Kafka 7.4.0, confluent-kafka 2.6.1 | | Database | PostgreSQL 18, SQLAlchemy 2.0.36, psycopg2 2.9.11 | | Feature Engineering | Python 3.13, Pandas 2.2.3 | | Machine Learning | scikit-learn 1.5.2, XGBoost 2.1.3, imbalanced-learn 0.12.4 | | Experiment Tracking | MLflow 2.18.0 | | API Se …