This talk presents the evolution of an enterprise MLOps platform from a custom AWS SageMaker-based registry, built around spoke-and-hub integrations, Lambda promotion workflows, canary deployments, and manual model governance, to a centralized Databricks and MLflow architecture. The transition was driven by the need to reduce maintenance overhead, improve onboarding, and standardize lifecycle management across LLM-based applications.

A key design principle is treating pipelines and prompts as versioned, reproducible artefacts. In this framework, prompts become first-class registry objects enriched with metadata, evaluation results, and promotion logic, enabling consistent governance from experimentation to production. The talk also covers practical LLMOps challenges, including structured outputs, fallback design, external scoring integration, and auditability in production environments.

This talk is geared toward Data scientists, ML engineers and MLOps engineers involved in building and scaling enterprise GenAI systems. Participants will learn how to evolve from a highly customized MLOps setup to a more scalable MLflow-based architecture, how to manage prompts and pipelines as versioned artefacts, and how to operationalize LLMOps practices such as evaluation, promotion, governance, and production resilience.