CIIR PhD alum Hema Raghavan will be on campus to present Kumo's innovative research, discuss her path to co-founding Kumo, and offer insights for aspiring entrepreneurs.
Speaker: Hema Raghavan, Co-founder, Kumo (NVIDIA)
Title: Replacing XGBoost: Open Foundation Models for Structured Data, and What It Took to Build Them
Date: November 23, 2026, 1:30 - 2:30 pm in CSL Rm. E120; reception immediately following the presentation
Abstract: For two decades, predictive modeling on enterprise data has followed one recipe: flatten a relational database into a single table and fit gradient-boosted trees such as XGBoost. Flattening discards the relational structure where much of the signal lives, such as customers linked to orders, orders to products, and transactions to accounts. To recover some of that signal, data scientists wrote hand-coded aggregations, such as a customer's spend over the last 30 days, the number of distinct products viewed, or the time since the last purchase. Accuracy came to depend almost entirely on how well a human could engineer features. It also made every new prediction its own project. In the years I spent building recommender systems at tech-forward companies, I saw hundreds of models running on thousands of feature pipelines. The pattern was the same everywhere: each model had to be built, validated, and maintained separately, so cost and engineering effort grew with every new task. Meanwhile, the field of text was changing. Practitioners were no longer engineering linguistic features such as n-grams, part-of-speech tags, or parse-based features. Models learned representations directly from raw text. That observation, together with our frustration at managing a multitude of pipelines, was the genesis of Kumo. We bet that a relational database is a graph (rows are nodes, primary/foreign keys are edges), so models should learn from that structure directly rather than from hand-engineered features. Graph neural networks proved the idea but still required one trained model per task, which led us to pretrained foundation models for structured data that learn from the database through in-context learning. This matters now because AI agents need many predictions on demand. Kumo is part of NVIDIA's Agent Knowledge Foundation, a set of open models and libraries for grounding agents in enterprise data. An agent states a predictive query, for example which customers will churn next month, and the model reads labeled historical examples from the database as context and returns a prediction without training a task-specific model.
I'll share my journey from a PhD at UMass Amherst, through roles at Yahoo!, IBM, and LinkedIn, to co-founding Kumo with Jure Leskovec and Vanja Josifovski. The NVIDIA Kumo family now covers Tabular, Relational, Forecast, and Anomaly models, plus Kumo Engine, an open-source accelerator for context generation and for training predictive models over large-scale structured and relational data. I'll also share the story of building the company, including the decisions we got right, the ones we got wrong, and the lessons I would pass on to researchers and students considering the leap from research to product.
Bio: Hema Raghavan is a co-founder of Kumo, now part of NVIDIA, where she builds foundation models for structured data. She holds a PhD in Computer Science from UMass Amherst (CIIR) and previously worked at Yahoo!, IBM Research, and LinkedIn, where she built large-scale machine learning systems. Kumo's work has been recognized with a place on the Forbes AI 50 list (2024), a Fast Company Next Big Things in Tech award (Foundational AI, 2025), and a place on Inc.'s 2025 Best in Business list (Best Startups). Hema was named to Inc.'s 2026 Female Founders 500. She is passionate about building inclusive organizations with a culture that makes innovation possible, because she believes the best ideas come from teams where people with different perspectives are able to contribute fully.
RSVP requested: Please fill out the RSVP form or email Ben Stickan at bstickan@uma-foundation.org to let us know if you will attend the presentation.
