AAIF x Uber: Inside Michelangelo
The platform behind every Uber ETA is now open source. Come see how it works.
Since 2016, Michelangelo has been the platform Uber's engineers use to build, train, deploy and monitor machine learning, from ETAs and pricing to fraud detection. Over time it grew from tree models to deep learning to LLMs, and today it also carries Uber's agent workloads. In May 2026, Uber open-sourced the core of it under Apache 2.0.
For this edition, the AAIF Community Amsterdam and Uber are spending an evening on Michelangelo. We'll cover how it's built, what running ML and agents at Uber's scale has taught the team, and what the open-source release means for your own platform. Expect architecture, trade-offs and real production numbers, not a product pitch.
What to expect
How Michelangelo is built: Kubernetes-native, Python-first pipelines with Uniflow, and federated compute across clouds
What it takes to serve ML at Uber's scale
How Uber runs LLMs and agents on top of its ML platform
A walkthrough of the open-source release and how to contribute
Drinks, food, and good company
Who should come
ML engineers, platform engineers, data scientists and technical leads who build or run ML and agent infrastructure, and anyone curious how a platform at this scale is designed and run.
All experience levels are welcome.
Schedule
[TBD: schedule]
Meet the speakers
[TBD: speakers and bios]
Before you come
Want a head start? The code is at michelangelo-ai/michelangelo, and the docs at michelangelo-ai.org.