MLOps Platform

Problem

A data-science department with no shared path to production — every model deployed differently, and data scientists needed direct GCP access to ship anything.

What was built

Architectural decisions I owned:

• Separation of model classes, so GenAI models deploy within the classic ML model space.
• Where the feature store sits in the platform architecture.
• The orchestration model — Cloud Composer outside the platform vs. at the Big Data boundary.
• A CLI abstraction layer for deployment, testing, and prediction without direct GCP access.

Outcome

• Adopted across the client’s entire data-science department as the standard for building and deploying ML models.
• All models run on the same interface.
• Same code for stage and prod environment - from testing to prod predictions.
• New model types are onboarded onto it incrementally.

Lessons and Challenges

• Managing classical ML and generative AI in one environment
• Managing metadata of models enabling datascientists to track backtestings, stage and production deployments, and predictions.
• Automatic building of dependencies and images using CI/CD pipelines — Azure DevOps.