Interview, Fireside Chat, Other
Turning Graph AI into ROI ft Kumo’s Hema Raghavan
- Kumo AI enables predictive queries on relational data (tables) by automatically abstracting the underlying graph structure, allowing users to write SQL-like "predictive queries" without manual graph construction.
- The platform operates as "AutoML on GPUs," distinguishing it from previous CPU-based AutoML tools by utilizing Graph Neural Networks (GNNs) to automatically learn features and model ensembles rather than relying on manual feature engineering.
- Hema Raghavan, co-founder and head of engineering, previously led the "People You May Know" initiative at LinkedIn, leveraging graph learning to drive core metrics like monthly active users and ad revenue at scale.
- Kumo claims to deliver business value within four-week proofs of concept (POC) for diverse sectors including healthcare (demand forecasting), fintech (fraud detection), and consumer platforms (churn prediction, recommendation).
- The system integrates directly with data warehouses like Snowflake and Databricks via native deployment models (Snowpark Container Services and Databricks apps), ensuring data residency and eliminating the need for complex data migration or security reviews.
- Unlike traditional graph learning which often requires PhD-level expertise, Kumo provides a "self-driving car" experience with an option to "drive stick," allowing advanced users to inspect and tweak neural network parameters under the hood.
- Kumo's architecture optimizes costs by storing edges in a compressed format on CPUs for inference while reserving GPUs exclusively for training and message passing operations.
- The platform addresses explainability requirements for regulated industries (insurance, healthcare) by tracing predictions to specific tables, columns, and instance-level features used by the model.
- Synergies with Large Language Models (LLMs) are realized by feeding Kumo's behavioral predictions into Retrieval Augmented Generation (RAG) pipelines to ground LLM outputs in personalized, historical user data.
- Adoption barriers exist for companies that have not yet established a structured data landscape; Kumo requires organized relational data in warehouses rather than unstructured spreadsheets to function effectively.
- Recent trends indicate a surge in graph learning, with over 50% of papers at the KDD data mining conference focusing on Graph Neural Networks.
- Hema advises aspiring AI engineers to prioritize foundational mathematics, specifically linear algebra and probability, over transient programming languages or frameworks.
- Kumo's culture focuses on "empowering people to do more than they think they can," resulting in high retention and a "no regrettable turnover" environment by hiring top talent to solve complex problems.