How Jumio built a real-time feature store on AWS

This article describes how Jumio built a real-time feature store on AWS to support low-latency fraud detection and identity verification.

• Streaming-first architecture uses Kinesis, Apache Flink, and SageMaker Feature Store for real-time feature serving
• Achieved 95th-percentile latency of 16.9ms, well below the 100ms SLA requirement
• Tiered storage strategy combines in-memory ElastiCache for hot data with standard store for cold data
• Parallel offline feature store uses Firehose, S3, and Iceberg tables for model training and analysis
• Centralized feature definitions replaced fragmented team-specific approaches, reducing manual deployment overhead
• Saved approximately $120,000 annually in operational costs compared to previous disparate systems

The architecture demonstrates best practices for building scalable, low-latency feature stores that support real-time ML inference while maintaining cost efficiency and operational simplicity.