What we're building
A streaming ML pipeline that scores every M-Pesa transaction passing through eRev (and, eventually, our connector SDK) for fraud risk. Features are derived from transaction graph (sender → receiver patterns), velocity (this MSISDN moved KES 200k in 3 minutes), and entity patterns we've extracted from past disputes.
Why now
Counties using eRev process ~3-5k transactions a day. Manual review surfaces the obvious problems but misses the layered ones. We have the data to learn the patterns; we just need to ship it carefully false positives are taxes on legitimate payers, false negatives are losses we eat.
Where we are
In pilot at one county; running in shadow-mode (scores predictions, doesn't block) while we build the evaluation harness. Targeting decision-quality parity with manual review before we ship to live decisions.
Getting it right
- Explainable scores every flagged transaction comes with the top 3 features that drove the score, in plain English.
- Per-county thresholds different counties have different risk tolerances; the model exposes the knob.
- Routine bias evaluation we test the model against demographic strata before each release.