America First Credit Union optimized their MLOps stack with the Hopsworks Feature Store

America First Credit Union has more than 1.2 million members and 120 offices worldwide and is one of the largest credit unions in the United States with over over $16 billion in assets. They use Hopsworks as an instrumental part of their AI pipeline to prevent ATM Fraud. 

AFCU has been able to achieve exceptional gains over their previous process for training models by adopting the Hopsworks Feature Store as a core part of their workflow. These optimizations involve both a conversion from Oracle SQL centric workflows to PySpark workflows, as well as the use of HopsFS rather than Oracle RDBMS.

The Aim: 

AFCU required a solution that could power their Enterprise AI services. One that could bring all operational and analytical data into their machine learning processes and avoid the pitfalls of BI platforms that aren’t adapted  for predictive workloads. 

• Faster development cycle for new features
• Flexible development environment that allows data scientists to test feature logic in Python
• Integration with the existing data warehouse and AFCU model development workflow
• A single platform that supports development and productionalization of analytical and predictive models
• Start with on-prem and migrate to the cloud as their cloud native capacity grew and their legacy on-prem requirements decreased

Why Hopsworks? 

• Testing is easier from within existing python testing frameworks
• Data science workflow improved across the board
• Reducing the complexity and increasing the readability of new features
• Better visibility into existing features that improves reusability across models and use cases compared to their SQL counterparts
• Ability to publish features for online/real-time use
• Able to ramp new employees in a few days instead of months
• Feature ‘Shopping cart’ allows users to share and reuse their work collaboratively

Results: 

• 3-4x Productivity Gain while vastly simplifying their machine learning codebase/pipelines
• Python-Centric Approach, removed the complexities with SQL
• ROI within 6 Months