The problem
Payment fraud in fast-moving ecosystems is difficult to catch with isolated signals. Labels are imbalanced, fraud patterns change quickly, and analysts need explanations before they can act on a score.
Loading page
Preparing the content for you.
Risk scoring that analysts can actually investigate.
A real-time fraud detection platform for African payment ecosystems. It combines calibrated gradient-boosting models, GraphSAGE graph learning, and deterministic rules to score transactions, explain risk, support analyst investigations, and identify coordinated fraud rings.
Overview
FraudSentinel scores payment transactions as they arrive, combining supervised models with graph-based signals and deterministic rules. The goal is not only to flag risk, but to explain enough of the decision for an analyst to investigate it.
Calibrated probabilities make scores comparable across transactions and time. Graph features surface relationships between accounts and devices, which helps identify coordinated activity that a single-transaction model would miss.
Payment fraud in fast-moving ecosystems is difficult to catch with isolated signals. Labels are imbalanced, fraud patterns change quickly, and analysts need explanations before they can act on a score.
FraudSentinel blends calibrated gradient-boosted models, graph learning, and explicit rules into one risk decision, then presents the contributing factors and related entities to an analyst for review.
Architecture
A high-level conceptual view of the main stages in this system.
Events arrive through a streaming pipeline and are validated before scoring.
Transaction features and relationship signals are assembled for each event.
Calibrated models and deterministic rules produce a comparable risk score.
Contributing factors and linked entities are presented to an analyst for a decision.
Capabilities that are documented as part of the current implementation.
Transactions are evaluated as they arrive rather than in overnight batches.
Gradient-boosting outputs are calibrated so scores can be compared and thresholded.
GraphSAGE features surface relationships that indicate coordinated activity.
Explicit policy rules cover known patterns and hard limits.
Each decision exposes the factors that contributed to the score.
Linked accounts, devices, and prior activity support a structured review.
These items are planned and are not available yet. They are listed separately so completed functionality is never confused with future work.
A reproducible evaluation report covering precision, recall, and calibration.
Connectors for more regional payment providers and settlement models.
A hosted sample dataset for walking through the analyst workflow.
Technology
The tools and technologies associated with the implementation.
Model training, feature engineering, and scoring services.
Gradient-boosted trees for transaction risk scoring.
Fast boosted models used for comparison and serving.
Graph model training and embedding generation.
Graph learning over accounts, devices, and transactions.
Low-latency scoring API for incoming transactions.
Service layer for analyst and administrative workflows.
Transaction, case, and review records.
Streaming transport for transaction events.
Challenges
The engineering constraints that shaped the design.
Fraud is a small minority of activity, so evaluation and training must handle rare positive labels without overfitting.
A score is only actionable if an analyst can see why it was produced and what evidence supports it.
Fraud behavior shifts, which requires monitoring, retraining, and rules that can respond quickly.
Outcomes
Qualitative outcomes from the design. Measured results are only reported where they have been formally evaluated.
Links
Explore the project through the destinations that exist. Missing destinations are labelled instead of being replaced with a placeholder URL.
Source code, issues, and technical documentation.
(opens in a new tab)No public link has been provided for this project yet.
No public link has been provided for this project yet.
No public link has been provided for this project yet.
1 of 4 external links are currently available. Unavailable links are shown deliberately rather than pointing to a placeholder destination.
Keep exploring
Browse the full portfolio or start a conversation about this project.