Temporal Context Graph For AI Agents

At Stripe, we process millions of API requests per second across 46 countries. Last quarter, our infrastructure team deployed inference models at the edge — moving fraud detection and payment routing closer to the merchant.

The results exceeded our projections. Median API response times dropped by 40%, with the largest gains in fraud detection where round-trips to centralized GPU clusters had been the primary bottleneck.

This post details the architecture, the tradeoffs we navigated, and what we learned about deploying ML models at the edge of a global payments network.

Why the edge?   Every millisecond matters in payments. When a customer taps their card at a terminal in Dublin, the fraud check previously had to round-trip to us-east-1 — adding 80-120ms of pure network latency. Multiply that by billions of transactions and you start to see why we were motivated.

Our approach was to deploy lightweight inference models to 14 edge locations globally, co-located with our existing API infrastructure. The models run on custom ONNX runtimes optimized for CPU inference — no GPUs required at the edge.

The fraud detection model was the first candidate. It evaluates 47 features per transaction: merchant category, velocity patterns, device fingerprint, geographic anomaly scores, and behavioral signals from the session. Previously this ran on a centralized cluster in Virginia.

After six weeks of shadow testing, we cut over Dublin, London, and Singapore. The p50 latency for fraud scoring dropped from 145ms to 12ms. The p99 dropped from 380ms to 34ms. We were stunned.