At the scale Uber operates, machine learning is the ground of the entire user experience. From matching riders with drivers to predicting hyper-local ETAs amidst rush-hour traffic, Uber’s infrastructure handles an astonishing 30 million predictions per second. By the time one has finished reading this sentence, millions of AI-driven decisions will have occurred worldwide.
To orchestrate this massive ecosystem, Uber relies on Michelangelo, its legendary, end-to-end internal ML platform. Michelangelo manages the entire model lifecycle: from feature selection and distributed training using Ray, to deep learning deployments on a fleet of over a million CPUs and GPUs.
But as AI becomes more deeply driven into society, a new challenge has emerged that goes far beyond raw computing power: Responsible AI and Governance.
The Governance Challenge: Enter the EU AI Act
With the arrival of the landmark EU AI Act, global tech giants have had to radically rethink how they deploy AI. The regulation enforces strict compliance on “high-risk” systems, which includes automated task allocation, like Uber’s core matching algorithms.
For an organization with thousands of engineers and thousands of microservices, verifying compliance is a very important task. The legal team could not realistically review millions of code files to determine which ones posed a regulatory risk.
To solve this, Uber’s ML platform team used AI to govern AI.
Scale, Generative AI, and “Model Cards”
Uber developed an automated, GenAI-assisted risk assessment system built directly into Michelangelo. By leveraging Large Language Models (LLMs) to scan repositories and notebooks, the system flags code files containing high-risk or prohibited indicators. Because legal teams don’t read raw Python code, the LLM also generates plain-English explanations detailing why a specific model might pose a compliance risk.
To solidify this process, Uber introduced Model Cards: a centralized registry where model owners self-certify their use cases.
By prioritizing recall over precision, the system ensures no potential risk goes unnoticed, saving over 100 years of manual legal review while establishing a future-proof foundation for upcoming global AI legislation.
Ensuring Fairness in the Grid
Beyond legal compliance, Uber is leaning heavily into algorithmic fairness. The platform now features integrated fairness estimators to automatically evaluate data and model drift, tracking metrics like false positive rates across different demographics. If an elite “Tier 1” model experiences unexpected bias or data drift in production, engineers are alerted instantly via PagerDuty and treating ML health with the same urgency as a critical backend outage.
This glimpse into Uber’s Michelangelo platform and their groundbreaking compliance framework is just the tip of the iceberg.
- How exactly does Uber optimize its 2D pricing and dispatch grids in real-time?
- What are the precise prompts and LLM frameworks (like LangFX) used to automate their legal triage?
- What did Uber’s ML platform team learn from their friction points with the legal department?
Gain access to the full video presentation and deep-dive technical breakdowns by subscribing to the hyperight.com platform. Want to network with the minds behind these systems? Hear directly from the engineering leaders shaping the future of scalable, responsible AI live at the upcoming summit.
