Manifold is an open-source visual debugging tool from Uber that helps machine learning practitioners understand where their models succeed and where they fail. With about 1,700 stars on GitHub, it is model-agnostic and stable, and it is being incubated for long-term support as part of Uber’s open-source program.
Summary metrics like AUC or RMSE hide the subsets of data where a model actually goes wrong. Manifold lets you look beyond those aggregates with interactive visual analytics: slice performance by data segments, compare models side by side, and inspect the feature distributions behind prediction errors. It works with structured data you feed it and renders an explorable dashboard, either through the demo app or embedded as a component in your own tooling. Because it is model-agnostic, the same workflow applies whether you are debugging a classifier, a regressor or a ranking model, which made it a standard part of Uber’s internal model development process before its open-source release.
Manifold is completely free and open source under the Apache-2.0 license. There is no hosted version or commercial tier; you run it on your own data and infrastructure. It is a visualization library, so your only costs are the environment you choose to deploy it in.
ML engineers and data scientists who need to diagnose model behavior beyond headline metrics, especially teams debugging performance gaps on specific customer or data segments. It rewards users who think visually about model errors rather than scrolling through logs and aggregate scores.