Skip to main content

Welcome to Michelangelo AI

Michelangelo AI is an end-to-end ML platform for building, deploying, and managing machine learning models. Born at Uber — where it powers 25,000+ model trainings per month and ~30 million predictions per second — now open source.

Why Michelangelo AI​

Most ML teams spend more engineering time on infrastructure than on models. Training pipelines break in production. Feature logic gets duplicated across teams. Serving containers are hand-rolled for each project. Stitching together a training framework, an experiment tracker, an orchestrator, and a custom serving layer works until it doesn't — and when it breaks, it's your problem to fix.

Michelangelo AI exists so you don't have to build that stack from scratch. It covers the full ML lifecycle — data preparation, training, evaluation, deployment, and monitoring — in a single system that has been running in production at Uber's scale for eight years. A complete platform, battle-tested across thousands of models, now available for any team to adopt.

Michelangelo AI is worth considering if:

  • Your ML team is spending more time on infrastructure than on models
  • You're managing a patchwork of tools with brittle integrations between them
  • You want a platform that handles classic ML, deep learning, and generative AI workflows without switching systems
  • You're building or inheriting an ML platform and want a proven foundation rather than starting from scratch
  • You want to contribute to — or build on — an open-source ML platform with real production history behind it

The platform is modular: adopt the full stack or integrate individual components (feature store, model registry, serving) into your existing infrastructure. It's designed to be extended, not to lock you in.

Get started​

Prefer an ordered path over a menu? Learning Paths sequences the docs into three role-based tracks (ML Engineer, Operator, Contributor) with time estimates.

I'm evaluating Michelangelo AI​

Understand what the platform does, how it compares to your current stack, and whether it fits your use case.

  • Overview — What Michelangelo AI is, how it works, and how familiar tools map to it
  • Core Concepts — Projects, workflows, tasks, and the key terms you'll encounter
  • Roadmap — Release milestones, what's available now, and what's planned

I want to build my first pipeline​

Get a local environment running and build an end-to-end ML pipeline.

I'm deploying or operating the platform​

Set up infrastructure, configure compute clusters, and deploy the UI.

I want to contribute​

What Michelangelo AI is — and isn't​

Understanding scope helps you decide if Michelangelo AI is the right tool.

Michelangelo AI is:

  • An ML lifecycle platform — data prep, training, evaluation, deployment, and monitoring in one system
  • An orchestration framework (Uniflow) for writing ML pipelines as Python code with @task and @workflow decorators
  • A model registry for versioning, tracking, and managing trained models
  • A deployment system for online inference (Triton) and batch predictions
  • A no-code UI (MA Studio) for standard ML workflows without writing code

Michelangelo AI is not:

  • A notebook environment — use Jupyter/Colab for exploration, then bring your code to Michelangelo AI for production
  • A data warehouse — it connects to your existing data sources (S3, Snowflake, BigQuery, HDFS)
  • A general-purpose workflow engine — it's purpose-built for ML, not arbitrary DAGs
  • A model monitoring SaaS — monitoring is built in, but Michelangelo AI is self-hosted infrastructure, not a managed service
  • A replacement for your ML framework — use PyTorch, TensorFlow, XGBoost, scikit-learn as you normally would

Support & Community​

Connect with the Michelangelo AI core team and developer community through our dedicated channels: