Michelangelo (Uber’s ML Platform): Going Open Source | Sally Lee & Eric Wang, Uber

Michelangelo (Uber’s ML Platform): Going Open Source | Sally Lee & Eric Wang, Uber

by Sally Lee & Eric Wang, Uber

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 Sally Lee & Eric Wang, Uber 👥 5K 📅 October 23, 2025 ⏱ 32 min 👁 978 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

MichelangeloMLOpsopen sourceUberML platform

Summary

In this talk at MLOps World 2025, Sally Lee and Eric Wang from Uber present Michelangelo, Uber’s end-to-end machine learning platform, and announce its upcoming open-source release. They describe the platform’s evolution over the past decade, from classical ML to deep learning and now to a multi-agent platform. Michelangelo supports over 25,000 model trainings per month and serves 6,000 models in production, handling 30 million predictions per second. The platform is built on Uber’s infrastructure and supports multi-cloud deployments. The talk includes a live demo by Eric Wang, illustrating the data scientist and ML engineer journeys. Key features highlighted include project scaffolding, a Python-first workflow engine called Uniflow, model registry, evaluation pipelines, retraining processes, and safe deployment strategies like Jono rollout. The speakers emphasize the platform’s plug-in architecture, which allows integration with external systems. They announce that the first components, including Uniflow and the control plane, will be open-sourced later this year, with the full platform expected to be public early next year. They are seeking partnerships for early adoption and collaboration.

172 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the architecture and operational practices of a large-scale ML platform. The speakers demonstrate a deep understanding of MLOps challenges and present concrete solutions, such as the Uniflow workflow engine and the plug-in system. The argumentation is solid, supported by real-world examples and a live demo. However, the presentation is high-level and lacks detailed technical depth, which may limit its value for practitioners seeking implementation details.

Scientific Rigor, Source Quality, Title Accuracy

The presentation is based on the speakers’ direct experience at Uber, lending it credibility. However, no external sources or published papers are cited, and the claims about scale and performance are not independently verified. The title accurately reflects the content, and the talk is well-structured. The live demo adds authenticity, but the lack of references reduces the overall scientific rigor.

146 words

Title / Content Match

Title accurately reflects the content: presentation of Michelangelo's architecture and open-source plans.

Quality & Reliability

7/10

Presentation by senior Uber engineers with concrete architectural details and live demo, but limited external validation and no published benchmarks.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • MLOps World — Conference where the talk was recorded.

Concurring Sources

  • MLOps World — Conference context aligns with the talk's theme.

Contribution & Novelties

The talk provides an inside look at Uber’s ML platform and its open-source plans, which is valuable for the MLOps community. The Uniflow workflow engine and the plug-in architecture are notable contributions. However, the presentation is more of an overview than a detailed technical guide.

Pour aller plus loin :

  • MLOps — Overview of MLOps practices.
  • Kubeflow — An open-source ML platform with similar goals.
  • MLflow — Open-source platform for ML lifecycle management.

73 words

Radar Profile

The radar profile shows high scores in information quantity and technical level, but slightly lower in reliability due to lack of external references. The overall balance suggests a technically informative but not fully rigorous presentation.

Reliability 7/10

💬 No comments were provided for analysis.