Scaling AI for high-stakes, real-time payments
summary
Stripe processes $1.4 trillion in 2024 and scales real‑time ML by building a Kron‑based feature store called Shepherd (two years to build), moving heavy inference out of the Ruby monolith into a sidecar/warehouse library, and parallelizing with an optimistic kick off early resolve later pattern that keeps common good payments fast while rolling back the rare fraud cases. Measured wins: unnecessary challenges down 20%, fraud down 8%, $6B recovered, and API reliability at 99.99986%.
AI-written summary · about
chapters (8)
- 0:00Scaling AI in payments
- 2:57AI in payment lifecycle
- 6:26Data quality challenges
- 8:52Measuring AI impact
- 12:21Reliable AI at scale
- 14:53Latency in real-time AI
- 21:18Collaboration and changes
- 24:50Key takeaways and future
related talks
15mScaling AI-Driven Financial Services: Ethical ML at Scale
32mMachine Learning Infrastructure at Stripe: Bridging from Python JVM
37mThe Role of Machine Learning in Today’s Payment Ecosystems
07mKeynote: Engineering Population-Scale AI & AI Infrastructure with Cloud Native Tec...
13mScaling AI Systems for High-Volume Transaction Processing
06mArtificial Intelligence - the “refinery” for Data
26mAI-Powered Fraud Detection in Cloud-Native Digital Payments
39mTrain, Evaluate, Repeat: Building a Credit Card Fraud Detection System