Jumio's AWS feature store delivers sub-100ms fraud detection

Jumio, an identity verification provider, tackled the challenges of real-time fraud detection by building a centralized feature store on AWS. Previously, fragmented feature engineering caused data duplication, manual production re-implementation, and latency issues. To fix this, they deployed a streaming-first architecture across us-east-1, eu-central-1, and ap-southeast-1, ensuring sub-100ms feature serving for ML inference. The design relies on Amazon Kinesis Data Streams to ingest events, with Apache Flink processing and enriching them in-flight before writing directly to Amazon SageMaker Feature Store. This real-time path handles immediate fraud-detection needs, including conditional feature creation based on event times.

For model retraining and offline analysis, a parallel pipeline uses Amazon Data Firehose to deliver raw events to Amazon S3, where event notifications trigger Amazon EMR for heavy transformations, landing processed features as Iceberg tables. This dual-path approach ensures consistent feature definitions across online and offline stores, eliminating redundant data and manual reimplementation. Jumio's architecture also scales to a growing feature catalog with schema evolution, while supporting delayed event patterns that arrive weeks later due to extended reviews. The result is a reusable, real-time feature store that meets the demands of identity verification—enabling businesses to detect fraud faster and build digital trust with confidence.