Learning from Extremes: What Fraud-Fighting at Scale Can Teach Us About MLOps Across Domains

The engineers behind large-scale anti-fraud platforms have been pioneers in MLOps due to extreme demands for low-latency inference, feature freshness, and agile redeployment. This talk challenges the assumption that these advanced architectures are overkill for less demanding domains. Instead, it argues that a real-time-first approach simplifies architectures by eliminating complex pipelines and that the observability and replay technologies developed for fraud can make ML teams more agile across the board. The presentation covers five dimensions of challenge in anti-fraud ML pipelines and discusses architectural approaches to solve them, drawing parallels to real-time ML in general.