Kumo (Kumo.AI) is an American AI company that develops models that make predictions directly on enterprise relational databases and tabular data. It was founded in 2021 by Jure Leskovec, a Stanford University professor and graph neural network researcher, together with Vanja Josifovski and Hema Raghavan, and is headquartered in Mountain View, California.
Its approach treats data spread across related tables as a graph and trains graph neural networks on it. In 2025 the company released KumoRFM, a relational foundation model that produces predictions on relational data without task-specific training. The goal is to handle prediction tasks such as churn prediction, fraud detection, and recommendation—traditionally built as separate pipelines with tools like XGBoost—with a single model.