Capital One

Solution

Capital One migrated its anomaly-detection machine learning workload from a legacy container-oriented design to a serverless implementation on Amazon Web Services (AWS), centered on AWS Lambda. Transaction-volume data from Capital One’s digital platforms, including mobile-app and web-channel critical transactions, is prepared for model training and inference in smaller components rather than one large monolith. To make the workload fit serverless constraints, the team sampled training datasets to cut data volume while preserving model performance, then packaged ML dependencies with Lambda layers and, where necessary, compressed and dynamically installed additional libraries at runtime. Capital One also reviewed any function exceeding Lambda’s runtime ceiling and refactored code or used an AWS asynchronous ML inference option for longer-running steps. Databases on the AWS host automatically scale with demand, while model components are deployed as independently managed microservices to improve scalability, reduce maintenance effort, and speed remediation of customer-impacting anomalies.

Data flow

Volume data from Capital One's mobile-app and web critical transactions is collected in AWS, sampled to reduce training size, and processed by anomaly-detection functions running in AWS Lambda. Those functions load ML libraries through Lambda layers or compressed runtime installs, interact with automatically scaling databases, and return anomaly signals so teams can investigate and fix customer-impacting transaction issues quickly.