Stripe details scaling Radar: fuller training history plus agent-assisted ML cut fraud
emilygsands · x · 2026-10-20
Stripe's full blog post on scaling Radar: the fraud models lean on network-wide signal connections (a new card at one business may have years of trusted history elsewhere on Stripe; an IP links activity across cards). Shield scores payments in milliseconds.
Motivation: despite 800M training transactions, Shield used only a fraction of available history — most payments with no reported fraud were discarded for efficiency. Fuller history should help separate fraud from behavior that's unusual network-wide but normal for a specific card or business.
Costs: a single pass took a day, recent payments had incomplete fraud labels, and more experiments outpaced manual capacity — hence the agent-assisted experiment loop. The winning SWA setup now prevents tens of millions in additional fraud per year.
More from coding & agent
- Runway Launches Grok Bot Plugin for Cursor, Bringing Multi-Model Media Generation to Coding Workflow — tlakomy · 2026-10-20
- esp32sim-mcp: MCP server lets AI agents build and test ESP32 firmware in a local emulator — EmanueleStrazzullo · 2026-10-20
- GameSummoner one-line local install: prompt-to-game pipeline runs fully local on a 5090 — demonicpigg · 2026-10-20
- Heavy Grok Bot user argues personal agents must stay open-source, shares backup SOP — WolframRvnwlf · 2026-10-20
- The ideal AI coding flow is 2-3 day immersive sprints, not 9-to-5 office hours — DesireeCachette · 2026-10-20
- Recall.ai publishes full build for an AI voice agent that joins Zoom, Meet or Teams calls — AssemblyAI · 2026-10-20