arikos/uganda-fraud-app

End-to-end Machine Learning pipeline & CLI inference engine for mobile money fraud detection in degraded financial ledgers. # 🛡️ Production-Grade Mobile Money Fraud Detection Engine > An educational, end-to-end Machine Learning case study and CLI inference utility for identifying transactional fraud within localized mobile money ledgers (e.g., MTN Mobile Money, Airtel Money). --- ## 📌 Architectural Flow & Execution Pipeline ```text ┌─────────────────────────────────────────────────────────┐ │ Corrupted Raw Synthetic Transaction Ledger │ │ (Negative balances, duplicate offline retries, dirty types) │ └────────────────────────────┬────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Phase 1: Data Cleaning │ │ • Normalize types (TRANSFER, CASH_OUT, PAYMENT, DEBIT) │ │ • Correct negative sign flips via abs() balance mapping │ │ • Deduplicate offline network retries │ └────────────────────────────┬────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Phase 2: Feature Engineering │ │ • Calculate errorBalanceOrig & errorBalanceDest │ │ • Flag isFullDrain (Account Takeover / ATO) │ │ • Track step-level transaction velocity │ └────────────────────────────┬────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Phase 3 & 4: Model Training & Evaluation │ │ • Stratified train/test splitting │ │ • Class-weighted Random Forest classifier │ │ • Metric analysis: ROC-AUC, Recall, F1-Score │ └────────────────────────────┬────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Phase 5 & 6: Export & CLI Inference │ │ • Model binary serialization (.joblib) │ │ • Low-latency CLI transaction risk scoring engine │ └─────────────────────────────────────────────────────────┘