Sentinel: AI Fraud Detection & Sanctions Screening

The results

55% cost reduction
Cost-optimized decision thresholds vs. baseline

sub-50ms scoring latency
Real-time fraud scoring with on-demand SHAP explainability

97.5% sanctions precision
Screen against the US Treasury's Office of Foreign Assets Control (OFAC) lists with confidence

ROC-AUC 0.8861
Exceeds the 0.85 industry benchmark for fraud detection

The problem

Payment companies are caught between competing pressures. Card-not-present fraud keeps climbing, and it costs cross-border businesses around 1.5% of transaction volume, which works out to $1.5M lost on every $100M processed. The defenses meant to catch that fraud create a problem of their own, because legacy systems flag 10 to 20% of good customers and steadily erode conversion. At the same time, OFAC sanctions screening has to resolve in milliseconds, and the rule-based systems built for an earlier era struggle to keep pace with fraud patterns that shift from week to week.

The solution

Sentinel combines ML-based fraud detection with real-time sanctions screening:

Fraud detection

LightGBM gradient boosting trained on 429 features
Explainable AI (SHAP values) for regulatory compliance
Calibrated probabilities with cost-optimized thresholds

Sanctions screening

Fuzzy name matching against 39,350 OFAC entities
Handles typos, name variations, and transliterations
Sub-50ms screening with 97.5% precision

Production-ready API

FastAPI with async processing
Redis caching for sub-200ms latency
PostgreSQL audit logs for compliance

How it works

Simple process

Submit a transaction via REST API or demo UI
Get a fraud score in under 50ms, with on-demand SHAP explainability
Accept, decline, or flag for manual review with a full audit trail

Technical implementation

Built with LightGBM for fraud detection, fuzzy matching for sanctions screening, and FastAPI for real-time API performance. Includes:

Instant results: fraud risk score, SHAP explanations, sanctions match
Transparent decisions: auditable predictions for compliance teams
Production stack: Redis caching, PostgreSQL audit logs, Next.js dashboard

Try live demo | Talk to us

Research project disclaimer

WARNING

This is a research case study demonstrating production-grade ML engineering. Models are trained on public datasets (IEEE-CIS, PaySim) for non-commercial research use only.

For production deployments: We train custom models on your transaction data, achieving 20-30% higher accuracy by learning your specific fraud patterns.

Try it yourself

Try the live demo to run it against sample transactions, or view the source code, which is Apache 2.0 licensed.

About Devbrew

Devbrew is a software agency. We design and build custom software and AI products for founders and growing companies, from the first prototype through to production.

Questions? Contact us | View repo: GitHub | LinkedIn