Kai Wang says Michelangelo supports Uber's business-critical machine learning and AI use cases across the full model lifecycle.
2
Uber gives most developers pre-built templates while allowing advanced users to access lower-level infrastructure and create customized workflows.
3
Uber is extending Michelangelo from a model platform into an agent platform, with plans to open-source parts of its orchestration framework.
Summary
Kai Wang explains how Uber built Michelangelo after teams began creating separate, hard-to-manage machine learning workflows around 2015. The platform now provides templates and pipelines for most developers, while advanced users can access infrastructure directly through tools such as UniFlow and Ray. Uber measures platform use through project feedback and proxy metrics because time to production varies widely by model and team. Wang describes how machine learning supports core Uber flows such as fraud detection, search, matching, pricing, ETA, recommendations, and customer support. Generative AI is being used for Uber Eats personalization, menu descriptions, restaurant feedback summaries, fraud detection, background checks, and internal data queries. Michelangelo is being extended with an agent studio, registries, MCP support, and permission controls. Uber plans to open-source parts of the platform, starting with UniFlow, after working with enterprise partners.
Michelangelo covers the full range of AI used at Uber
Kai Wang separates AI from the narrow public image of ChatGPT. At Uber, AI includes simple linear models, tree-based models such as random forests and XGBoost, traditional deep learning, and generative models. Uber's machine learning journey began around 2015, when teams working on maps, pricing, and risk replaced rule-based systems with machine learning. Those teams initially built separate Python workflows in notebooks. The code was difficult to share, reuse, productionize, and scale, which led Uber to create Michelangelo as a centralized platform for the machine learning lifecycle.
Developers need both simple defaults and access to lower-level infrastructure
Kai says Uber learned to let developers choose their tools. Michelangelo provides an abstraction layer, pre-built templates, and pipelines for more than 80% of users who want to build applications quickly. The remaining advanced users need customized workflows, different trainers, and direct infrastructure access. UniFlow, described as a Python orchestration framework, lets those users write their own code, run training on Ray clusters, and deploy models through Michelangelo. This split gives beginners a simpler path without preventing experienced teams from changing the underlying workflow.
Developer velocity is measured through project context and proxy metrics
Michelangelo's product goal is to help developers rapidly build and iterate on machine learning applications. Kai identifies time to production as the ideal measure, from project idea to production launch, but says it is difficult to compare because projects and team capabilities vary. A linear model and a model trained from scratch have different timelines, as do experienced engineering teams and newly graduated applied scientists. Uber therefore combines team feedback with platform-level indicators such as models trained, models deployed, evaluation pipelines run, and reports generated. One pricing team reported an 80% reduction in engineering cycles after using Michelangelo.
After most Uber teams had adopted machine learning in their core flows by late 2019 or early 2020, Michelangelo shifted from enabling adoption to improving project performance. Uber introduced four tiers based mainly on business impact. Tier one includes pricing, rider-driver matching, ETA, and fraud detection, where model failure can cause a high-level service outage. Tier four contains personal experiments. Kai says Uber has about 40 tier-one projects, about 100 tier-two projects, and roughly 500 to 600 tier-four projects. Tier-one projects receive attention first, and all of Uber's tier-one projects are currently predictive machine learning rather than generative AI.
Generative AI is being applied where users or internal teams can see the result
Kai describes several generative AI projects in Uber Eats. Large models help create more useful personalization by understanding menus, dishes, and user interests. They also improve search when users express concepts such as wanting something fancy, and they build a taxonomy for dishes and restaurants. Uber uses the resulting categories to generate more specific carousel titles and match restaurants to eaters. Another project summarizes restaurant reviews, scores, and other data for merchants, who can use the feedback to improve dishes or service. Across Uber, generative AI is also grouped into user experiences, process automation, and employee productivity.
Michelangelo has to support very different inference and compute needs
Extending Michelangelo to generative AI requires more high-end GPU capacity and changes to the technology stack. Large models often do not fit on one GPU, so Uber uses model parallelism. Kai says the platform integrated DeepSpeed and uses it with Ray for large-model fine-tuning. Uber also enabled NVIDIA Triton for serving large models. These additions help some traditional deep learning systems as well, including recommendation models, although smaller tree-based models do not need the same infrastructure. Michelangelo handles infrastructure-side optimization, while model teams handle choices such as quantization, so serving improvements require collaboration.
Uber is building an agent studio above its model studio
Michelangelo previously focused on training models, deploying endpoints, and making predictions. Uber is now building an agent studio for creating, evaluating, deploying, and managing AI agents at the application level. The studio will sit above model studio and use models built there. Uber is also working on an agent registry so teams can reuse agents created elsewhere instead of starting from scratch. Kai says permission management is part of the design because agents may access internal data and tools. The security team is helping define controls based on the user's or team's authentication and permissions.
Agent evaluation needs more than standard predictive metrics
Kai contrasts predictive machine learning evaluation, where teams can use metrics such as AUC, precision, and recall, with generative AI evaluation, which is still unsettled. Uber has built a generative AI evaluation framework with several approaches. One uses another language model as a judge of the output. Another includes humans in the evaluation loop. Teams can also provide a golden dataset and use it to assess performance. Kai stresses that speed alone is insufficient for systems such as text-to-SQL because the output must be correct.
Uber plans to open-source Michelangelo components after partner testing
Kai says Uber plans to open-source parts of Michelangelo, beginning with its orchestration framework, UniFlow. The team needs to clean the codebase first. During 2025, Uber planned to work with selected enterprise partners in a closed-source arrangement so they could access and contribute to the project. The broader opening was planned for the following year. Kai gives two reasons for the change: the machine learning tooling ecosystem is much larger than when Uber started Michelangelo in 2016, and generative AI is advancing faster than a 100-person team can track alone. External contributions could help the platform keep pace.
"AI actually covers the whole spectrum of machine learning from the very simple linear models to tree-based models like random forest and XGBoost to the traditional deep learning models like convolutional neural network and recurrent neural networks and all the way to generative AI."Kai Wang01:14
Who should watch
You run an internal ML platform and need a practical way to support both standard workflows and advanced custom projects.
Your organization is deciding how to prioritize predictive ML and generative AI projects by business impact and infrastructure cost.
You are designing agent platforms and need ideas for registries, permissions, evaluation, and reuse across teams.
