The Story

How is it that several years into the AI revolution, AI models still don't understand relational data?

The most useful business data — such as transactions, customers, and inventory — lives in your data warehouse. Yet AI models have historically struggled with relational structures. So while text-based applications have been radically improved with language models, predictive applications on top of relational data still require data scientists and engineers to spend months on feature engineering or lossy workarounds. Organizations spend millions per year just for incremental improvements because these applications are at the core of their business, but everyone from data scientists to engineers to executives wish they'd had more to show for their effort.

They can get their wish, thanks to two recent breakthroughs: Graph Neural Networks (GNN) and Relational Deep Learning (RDL). It's now possible to train accurate predictive and embedding models directly on relational data, without the need for feature engineering, extensive data prep, or data transformation. Data scientists and engineers can use it to train more accurate models in up to 95% less time than traditional ML methods or LLM-based workarounds.

Doing this in production at scale, however, comes with challenges. And nobody knows how to do it better than Jure Leskovec, who led the development of RDL and GNN at Stanford and then applied it as Chief Scientist at Pinterest; Vanja Josifovski, who implemented and scaled the technology as CTO of Pinterest and then the CTO of Airbnb; and Hema Raghavan did the same as an engineering leader at LinkedIn. They had a shared vision of helping teams get all the benefits of these new models without the cost, complexity, and expertise it once required — so they made Kumo.

Kumo is the platform for anyone to train and run state-of-the-art AI models on their relational data. Kumo models already power fraud detection, recommender systems, risk scoring, entity resolution, RAG, and other mission-critical predictive tasks inside applications used by hundreds of millions of people worldwide. Most importantly, it lets data scientists and engineers do more of what they love and make a greater impact all across their business.