Fintech · Machine learning
Hybrid GNN Fraud Intelligence
A team project for detecting fraud in mobile-money transactions using graph neural networks, XGBoost, and a FastAPI backend.
Context
Transaction-only checks can miss patterns that emerge across connected accounts, devices, agents, and institutions. This includes fraud rings, mule accounts, SIM-swap activity, reversal scams, and fast cash-out patterns.
Detection approach
XGBoost baseline
Uses transaction amount, frequency, recipient counts, linked accounts, shared devices, and time-based signals for fast tabular fraud scoring.
Graph neural network
Uses transaction-network structure to identify connected and organised patterns that are difficult to identify from individual transactions alone.
Stacked hybrid model
Combines graph embeddings with tabular features to provide a model that accounts for both behavioural and network signals.
Analyst workflow
The React interface includes live dashboard metrics, a transaction simulator, network visualisation, model comparison, alert-management actions, file upload, and explanations for flagged transactions.
Project snapshot
Role
Team project contributor on a product that connects machine-learning detection with an analyst-facing web experience.
Challenge
Detecting coordinated fraud that is difficult to surface through individual transaction features alone, while keeping findings understandable to an analyst.
Outcome
Delivered a working prototype that combines tabular and graph-based detection with dashboards, alerts, simulation, network visualisation, and explanations.
Technology
Python · FastAPI · PyTorch · PyTorch Geometric · XGBoost · Neo4j · React · TypeScript · Tailwind CSS · Docker