RiskGuard
AI-powered fraud detection and real-time risk scoring for card acquiring pipelines. Combines ML anomaly detection with rule-based decisioning.
Payment fraud is an arms race. Rule-based systems catch known fraud patterns — and fraudsters adapt. The velocity of new fraud techniques has outpaced the velocity of rule updates for most acquiring operations.
RiskGuard is a risk decisioning engine built for card acquiring that combines ML-based anomaly detection with interpretable rule-based decisioning, designed to run in the authorization path with sub-50ms latency requirements.
The Problem with Pure ML
ML models are excellent at detecting anomalies. They’re poor at explaining decisions — which creates two problems in payment risk:
Regulatory explainability. In many markets, adverse action decisions (declining a transaction) require an explainable reason. “The model said so” isn’t compliant.
Operator trust and tuning. Risk teams need to understand why transactions are being declined to tune thresholds, investigate patterns, and respond to merchant escalations. Black-box decisions make this impossible.
RiskGuard uses ML for what it’s good at (anomaly scoring, pattern detection across high-dimensional feature space) and rule-based logic for what requires explainability (final decisioning with auditable reason codes).
Architecture
Feature pipeline — real-time feature computation from transaction data, merchant history, cardholder behavior (where available), device signals, and network-level features. Feature computation runs in <5ms.
ML scoring layer — an ensemble of gradient boosting and neural network models producing a fraud probability score. Models are retrained weekly on labeled data. Separate models for card-present, card-not-present, and cross-border transaction types.
Rule engine — a configurable rule layer that takes the ML score as an input along with other signals and produces an approve/decline/review decision with reason codes. Rules are version-controlled and changes go through approval workflows.
Feedback loop — chargebacks, disputes, and manual review outcomes feed back into model training. The system improves continuously from its own decisions.
Results
On our acquiring volume:
- 3x improvement in true fraud detection rate versus the previous rule-only system
- 40% reduction in false positive rate (legitimate transactions declined)
- Authorization latency impact: median +12ms, p99 +38ms
- Full audit trail for every decision, exportable for regulatory review