FraudGuard

What it is

FraudGuard scores payment transactions from 0 to 1000, lets low-risk ones pass, sends uncertain ones to a ranked analyst queue with plain-language reasons, and retrains from the analysts’ decisions. It includes governance and suspicious-activity-report workflows.

The problem

Payment processors, online shops and mobile-money operators face account takeovers, shared-device rings and pay-on-delivery abuse. Static rules alone block genuine customers, and manual review cannot keep up with the volume.

What I built

I worked on a team product, so these are my parts:

• Velocity, behavioural and shared-entity (device, IP, phone) features, feeding an XGBoost and Isolation Forest ensemble, with training and scoring using the same feature definitions.
• Work on the outbox pattern between PostgreSQL and Kafka, so ingestion and predictions stay consistent when services restart.
• Airflow workflows for retraining, model promotion, drift monitoring and compliance reporting, logged to MLflow.
• Parts of the analyst dashboard (triage, audit trails, light and dark themes) and the commercial-readiness documentation.

How it’s checked

I do not publish accuracy or performance figures for this product, because it is company property and I do not have figures I can cite here.

Status

This page describes my part in a product owned by my employer, 3Consulting. A public product site exists at the link above. I make no claim here about where the machine-learning or retraining parts are deployed. My own separate open-source fraud project is Moniguard.

My role

My part at 3Consulting Limited, Lagos (AI Assurance & Solutions): feature engineering, the ensemble scoring pipeline, the MLflow and Airflow workflows, and the analyst dashboard and governance reporting.

Stack

Python, XGBoost, scikit-learn, Pandas, Kafka, PostgreSQL, Redis, MinIO, Airflow, MLflow, Flask, Prometheus, Grafana and Docker Compose.