duncanmwirigi/Credit-Scoring-for-East-Africa

Credit scoring for M-Pesa, SACCO, and bank lending channels. The system combines channel-specific feature engineering, a calibrated ML model, scorecard conversion, and deterministic policy rules — the same layered pattern used in mature fintech and banking stacks. # East Africa Credit Scoring Engine An **unbanked-first** credit scoring platform for East Africa. Most applicants have no traditional bank account — scoring relies on **M-Pesa**, **phone data** (SMS, call log, device tech), and **statements**, with **bank**, **SACCO**, and **CRB** data used when available. It combines channel-specific feature engineering, a calibrated ML model, scorecard conversion (300–850), deterministic policy rules, **SHAP explainability**, loan limit assignment, and a **FastAPI REST service** with regulatory audit trails. ## Platform at a glance | Capability | Status | Entry point | |------------|--------|-------------| | Multi-channel scoring (unbanked / M-Pesa / SACCO / bank / mobile lender) | ✅ | `score.py`, `/score` | | Loan limit assignment (increase / decrease / maintain) | ✅ | `src/lending/limit_engine.py` | | Borrowing & repayment history (all channels) | ✅ | `lending_history` in API | | Unbanked-first scoring (`channel: unbanked`) | ✅ | M-Pesa + phone; bank optional | | Optional bank / SACCO / CRB enrichment | ✅ | `data_sources` flags | | Phone tech (OS, RAM, tier, network) | ✅ | `phone_data.device` | | Phone data (SMS, call log, contacts, apps) | ✅ | `phone_data`, `src/data/phone_data.py` | | ML training + evaluation (AUC, Gini, KS) | ✅ | `train.py` | | Calibrated probability of default (PD) | ✅ | `src/ml/trainer.py` | | Scorecard mapping (PDO methodology) | ✅ | `src/ml/scorer.py` | | Policy engine (CRB, DTI, channel rules) | ✅ | `src/policy/engine.py` | | SHAP feature explainability | ✅ | `src/ml/explainability.py` | | Regulatory audit trail (JSON) | ✅ | `assets/audit_trails/` | | FastAPI REST service | ✅ | `serve.py` → port 8000 | | OpenAPI / Swagger docs | ✅ | `/docs` | | Synthetic training data (no PII) | ✅ | `src/data/synthetic.py` | ## Table of contents - Install - Verify it works - How to use this package - Supported channels - Data sources - Architecture - End-to-end workflow - Project structure - Training (`train.py`) …