Jumio Develops Real-Time Feature Store on AWS for Fraud Detection

Jumio, a provider of identity verification solutions, faced challenges with data duplication, feature engineering, and latency in managing real-time features for fraud detection. To address these issues, Jumio implemented a scalable, reliable architecture on AWS, combining streaming data processing with centralized feature storage.

The system employs Amazon Kinesis Data Streams for real-time data ingestion, with Apache Flink processing the data to generate features. These features are stored in Amazon SageMaker Feature Store, optimized for low-latency access during inference. Parallel batch processing pipelines ingest data into Amazon S3 via Amazon Data Firehose, with Amazon EMR transforming and populating features into Iceberg tables for offline analysis.

This architecture supports both real-time and offline feature stores, enabling rapid model inference and comprehensive data analysis. Monitoring tools track latency and system health, ensuring performance standards are maintained.

Compared to previous decentralized approaches, this centralized system has significantly improved operational efficiency, automated deployment, and handled late-arriving data effectively. It has also achieved a cost saving of approximately $120,000 annually.

The implementation emphasizes principles such as streaming-first design, centralized feature definitions, tiered storage, and cross-team collaboration, which collectively enhance model development speed, accuracy, and operational agility.