KubeCon Amsterdam 2026: The Industrialization of ML - A Deep Dive into Uber’s AI Platform Architecture

Uber outlined how its ML platform evolved from scattered scripts into Michelangelo, a standardized “machine learning factory” for training, serving, and feature management.

The next step was a Kubernetes- and Ray-based architecture, built to better support scalable deep learning and LLM workloads, including GPU-heavy jobs.

Uber’s stack now spans offline training, online serving, and feature stores across systems like Spark, Hive, Kafka, and Flink, aiming for more consistent operations at hyperscale.

The presentation shows how Uber industrialized ML infrastructure to manage millions of workloads and very high prediction volumes, offering a model for large-enterprise AI platforms.