Turning Graph AI into ROI ft Kumo’s Hema Raghavan

Kumo co-founder Hema Raghavan explains how GPU-powered graph neural networks replace manual feature engineering for enterprise predictive AI on Snowflake and Databricks.

• Kumo eliminates feature engineering pipelines by learning all features automatically via graph neural networks on GPUs, not CPUs.
• LinkedIn’s graph neural network rollout took 4-5 years and many engineers; Kumo delivers comparable capability as a native Snowflake or Databricks app.
• Kumo invented a SQL-like Predictive Query Language so analysts can specify ML problems without knowing graph theory.
• Within a 4-week proof of concept, Kumo claims it has never failed to show measurable business value to a customer.
• Kumo is a poor fit only for companies still in spreadsheets or those that cannot yet measure AI value — not a scale requirement.
• GPU-only compute is reserved for training; edges are stored in a compressed proprietary graph engine on CPUs to keep costs low.
• Kumo predictions can feed RAG pipelines as personalization signals, complementing LLMs that lack behavioral data.
• Raghavan advises aspiring AI engineers to prioritize probability and linear algebra over chasing the latest frameworks.

2025-01-21 · Watch on YouTube