An engine built to be audited
A calibrated machine-learning pipeline that explains every score and can run inside your own infrastructure.
Under the hood
- Model
- Random Forest ensemble: 200 trees, max depth 8
- Calibration
- Isotonic regression (Pool Adjacent Violators)
- Explainability
- Kernel SHAP contributions and counterfactuals
- Inputs
- Mobile money, utility payments, airtime, device metadata
- Training
- Synthetic population from real mobile money distributions, then calibrated on real outcomes
- Deployment
- Hosted API, or Docker on your own servers
Tested before it touches a loan
Feature ablation
How much accuracy is lost when top features are removed.
Data sparsity
Performance when many input fields are missing.
Distribution shift
Whether it holds across regional behaviour.
SHAP stability
Whether explanations stay consistent across resampled populations.
Governance built in
Data sovereignty
Scoring can run in your environment. No borrower data leaves unless you choose a hosted deployment.
Regulatory alignment
Designed with CEMAC and regional data protection frameworks in mind from the start.
Explainable by default
Every score ships with feature-level reasons for lenders and borrowers.