What blocks does an ML platform need? Learning from Uber Michelangelo

Introduction

An ML model that runs well in a notebook is not an ML system.

Production ML needs much more: right data, reproducible training, safe deployment, stable prediction, post-deployment monitoring, and the ability to rollback when errors occur.

Uber Michelangelo is a classic case study to understand why big companies need ML platforms.

1. What problem does Michelangelo solve?

According to Uber Engineering, Michelangelo was designed to help teams build, deploy, and operate machine learning solutions at Uber’s scale.

It covers the end-to-end ML workflow:

• data management;
• training;
• evaluation;
• deployment;
• prediction;
• monitoring.

Simply put, Michelangelo attempts to turn ML from isolated projects into a common platform.

2. Why is an ML platform necessary?

When every team builds their own custom pipeline, problems emerge:

• training and serving data are inconsistent;
• models are hard to reproduce;
• manual deployment;
• not knowing if the model is drifting or not;
• no way to track prediction quality;
• every project has to reinvent the wheel.

An ML platform is born to reduce those repetitive parts.

3. What could a minimalist ML platform include?

For a small project, you don’t need to build Michelangelo. But you can learn its structure:

Data Source
→ Feature / Dataset Version
→ Training Pipeline
→ Evaluation Report
→ Model Registry
→ Model Serving API
→ Monitoring
→ Rollback

An extremely compact version for a portfolio:

PostgreSQL / CSV
→ Training script
→ Evaluation metrics
→ Saved model
→ FastAPI endpoint
→ Docker deployment
→ Basic logs + latency metrics

4. Example: lead scoring in a CRM

A CRM that wants to score leads might need:

• lead data;
• interaction history;
• deal results;
• feature extraction;
• model training;
• an API to predict scores;
• a dashboard to display scores;
• monitoring to see if the score is still useful.

If you just train a model and stop, the system hasn’t created much value. Value lies in the score entering the workflow: which lead to call first, which follow-up to send, which quote to prioritize.

5. Conclusion

Uber Michelangelo shows that an ML platform is not a luxury. It is a natural reaction when ML enters production and multiple teams use it.

For small projects, you don’t need to copy the whole thing. But you should learn the principle: ML must have a clear lifecycle, not just a notebook.

References

• Uber Engineering — Meet Michelangelo: Uber’s Machine Learning Platform: https://www.uber.com/us/en/blog/michelangelo-machine-learning-platform/
• Uber Engineering — Scaling Machine Learning at Uber with Michelangelo: https://www.uber.com/us/en/blog/scaling-michelangelo/