Real-time fraud detection
Catching money-mule networks in milliseconds, with reasons an investigator can check.
Gen AI Architect · team lead
Money-mule networks, chains of accounts recruited to receive and pass on stolen funds, were causing significant losses. Looked at one transaction at a time, mule activity can look ordinary, and investigators needed to move faster without giving up the ability to explain why an account was flagged.
How it works.
Real-time scoring with XGBoost
A gradient-boosted model scores activity as it happens, fast enough (4.25 ms) to sit in the transaction path rather than in an overnight batch.
Graph analytics for the network view
Accounts and money flows are analysed as a graph, so the platform can surface connected rings of accounts that individual transaction checks miss.
Cost-based thresholding
Alert thresholds are set from the business cost of a missed fraud versus a false alarm, not a default cut-off, so the investigation queue matches what the bank can act on.
SHAP explainability
Every score comes with SHAP feature attributions, giving investigators and reviewers the "why" behind each alert.
GigaChat SLM investigation agent
An agent built on a GigaChat small language model helps investigators work a case, working from the model’s score and its explanation.
The platform reached 93.5% PR-AUC at 4.25 ms latency and cut fraud losses by 76% compared with the baseline, while keeping every decision explainable.
My role
- Led the team that architected and built the platform end to end
- Designed how the model, graph analytics, explainability and agent fit together
Stack
- XGBoost
- Graph analytics
- GigaChat SLM
- SHAP
- Python