Scaling ML in production: how BBVA accelerated delivery with MLOps

This article describes how BBVA standardized ML delivery across four business units by building reusable MLOps templates on Amazon SageMaker, combining real production use cases with architectural design to balance standardization and flexibility.

• Developed reusable MLOps templates with modular training and inference pipelines supporting diverse ML workloads
• Implemented step-level caching, CPU-first compute strategy, and conditional model registration to optimize costs and execution time
• Extended templates to support GPU-accelerated deep learning and hybrid PySpark-Python distributed processing without fragmenting workflows
• Embedded governance, experiment tracking, and CI/CD directly into the ML lifecycle with MLflow integration and Model Registry automation
• Achieved 30-75% improvements in efficiency through standardization, step caching, and migration from notebook-based to modular pipelines
• Established best practices including ephemeral validation environments, per-step resource optimization, and governance by design

BBVA's standardized but flexible MLOps operating model accelerates development, improves cost efficiency, and strengthens governance while preserving the flexibility required by different business domains and technical patterns.