Vamsi Chitters describes how Brex transitioned from a manual fraud detection process to a system where they automatically detect fraudulent transactions in real-time. This system leverages a model trained on streaming log data to calculate risk scores for ACH and wire transactions. Chitter details the streaming log pipeline that makes log data available to data scientists in Snowflake, including the Aggregator, Transport, and Connector components. He also provides insight into the technical requirements for these components, the options that Brex considered, and the rationale behind their choices. He concludes with an overview of their stress testing system and how they used it to fix a configuration bug.