SKitavi/akiba-ai

AI-Powered Offline Credit Scoring for African SACCOs # AkibaAI AkibaAI is an offline-first credit-risk assessment demonstration for African SACCOs. It turns synthetic mobile-money transaction evidence into behavioral features, produces an XGBoost risk score, explains the score with SHAP, and lets a human loan officer record a separate decision. > AkibaAI is an educational MVP that uses synthetic data. It is not a production > lending system and must not be used to make real credit decisions. ## Start here Read the complete system guide for a plain-language tour of the problem, architecture, assessment workflow, model, explanations, database, analytics, security boundaries, deployment, testing, and limitations. Additional reference documents: - Synthetic data schema - Contabo VPS deployment and operations ## What the application demonstrates 1. Select a synthetic member or enter a member ID. 2. Provide transaction evidence from demo data, CSV, supported SMS text, or a receipt image processed with local OCR. 3. Validate and normalize the evidence into one canonical transaction format. 4. Aggregate the transactions into 32 behavioral features. 5. Run the bundled XGBoost model locally. 6. Explain the score with traceable SHAP contributions in English or Kiswahili. 7. Save the assessment to SQLite. 8. Record a separate human decision: approve, review, or decline. 9. Review persistent history and operational analytics. No cloud model, generative-AI API, or internet connection is required for the assessment pipeline. ## Run locally ```bash python -m venv .venv # Windows PowerShell .venv\Scripts\Activate.ps1 # Linux/macOS source .venv/bin/activate pip install -r requirements.txt streamlit run src/ui/app.py ``` Open `localhost`. To load 12 idempotent synthetic dashboard records: ```bash python -m src.storage.seed_dashboard_demo ``` You can also load or reset demo assessments from **Settings**. The default demo Settings key is `CMU#AB39`; set `SETTINGS_ACCESS_KEY` to override it. ## Run the tests …