Fraud Detection AI: Real-Time GPU Inference for Transaction Monitoring

Thirty-Seven Million Transactions Per Month

A UK payment processor handles 37 million card transactions per month across 12,000 merchant accounts. Their rule-based fraud detection system flags 2.8% of transactions for review, but only 0.4% of flagged transactions are genuinely fraudulent — a false positive rate that costs £280,000 annually in manual review labour and an estimated £1.2 million in declined legitimate transactions. Meanwhile, sophisticated fraud patterns that bypass static rules account for £890,000 in annual losses. The processor needs AI models that reduce false positives by 70% while catching the evolving fraud patterns that rules miss.

GPU-accelerated fraud detection scores each transaction in under 8ms using a graph neural network that analyses transaction patterns, merchant relationships, and cardholder behaviour simultaneously. The model processes the full transaction context — not just the current transaction, but the cardholder’s recent history, the merchant’s typical patterns, and cross-account velocity checks — returning a fraud probability score before the authorisation timeout. A dedicated GPU server provides the consistent sub-10ms latency required for real-time transaction decisioning on private infrastructure.

AI Architecture for Transaction Monitoring

The system operates in two tiers. The real-time tier scores every transaction at authorisation time: a gradient-boosted model evaluates 150+ engineered features (transaction amount relative to cardholder average, time since last transaction, geographic velocity, merchant category risk score), while a graph neural network evaluates the transaction within the broader network of cardholder-merchant-device relationships. Transactions scoring above the threshold trigger immediate decline or step-up authentication.

The batch tier runs hourly, using an LLM to analyse flagged transaction clusters and generate natural-language explanations for the investigation team. The LLM summarises why a group of transactions was flagged, identifies the specific behavioural anomalies, and suggests investigation priorities.

GPU Requirements for Fraud Detection

The payment processor handling 37 million monthly transactions (approximately 14 transactions per second at peak) operates well within the RTX 5090’s capacity, with significant headroom for growth and model complexity increases.

Recommended Software Stack

• Real-Time Scoring: XGBoost with GPU acceleration for feature-based scoring (sub-2ms)
• Graph Analysis: PyTorch Geometric for cardholder-merchant relationship patterns
• Feature Store: Redis for real-time feature retrieval (cardholder profiles, velocity counters)
• Investigation AI: Llama 3 8B for generating natural-language case summaries
• Model Serving: NVIDIA Triton Inference Server for multi-model orchestration
• Monitoring: Prometheus and Grafana for latency, throughput, and model drift tracking

Regulatory Compliance and Model Governance

The Payment Services Regulations and FCA expectations require firms to maintain adequate fraud prevention systems. AI models used for transaction decisioning must be explainable — each decline needs a reason code the cardholder can understand. Model performance must be monitored for drift, with regular retraining on confirmed fraud labels. A GDPR-compliant dedicated server ensures transaction data, cardholder profiles, and fraud model weights remain within UK-controlled infrastructure with complete audit trails.

Getting Started

Extract 6 months of labelled transaction data (confirmed fraud and legitimate). Train the gradient-boosted model on 80% and validate on 20%, targeting an AUC above 0.97 and a false positive rate below 0.8%. Deploy in shadow mode: the model scores every transaction but does not influence decisioning for 30 days. Compare AI model flags against the existing rule-based system to quantify improvement. Organisations also running customer support AI or document processing can share the GPU server during off-peak hours. Browse additional finance use cases.