Interview, Fireside Chat, Other
Turning Graph AI into ROI ft Kumo’s Hema Raghavan
- Kumo enables immediate predictive query execution on relational data via graph abstraction and a proprietary query language resembling SQL with a "predict clause," while offering neural network parameter tuning for advanced users.
- The platform targets enterprise adoption across FinTech, delivery services, and healthcare by replacing manual feature engineering with Graph Neural Networks (GNNs) that learn all necessary features, potentially shifting standard time windows from 30 days to 365 days for churn prediction.
- Specific use cases include demand forecasting to stock emergency rooms, fraud detection for suspicious user behavior, and metrics prediction for video views or next-best-action recommendations.
- Kumo anticipates a four-week proof of concept that will almost always demonstrate value, with deployment as a native app in Snowflake and Databricks to reduce security friction by preventing data from leaving the ecosystem.
- Cost efficiency is planned through proprietary compressed edge storage on CPUs, reserving GPUs exclusively for training and message passing, despite GPU execution being significantly faster for scaling than CPU-only operations.
- The technology aims to integrate with Large Language Models by using GNN-inferred semantic representations to ground RAG algorithms and reduce hallucinations, while developing instance-level explainable AI based on specific table features.
- Market expectations indicate GNNs will become central to the AI revolution, evidenced by over half of KDD conference papers focusing on the technology, with potential application in chat agents combined with LangChain and Pinecone.
- The platform is expected to be unsuitable for organizations still relying on spreadsheets or in the early stages of warehouse migration, as it requires an established data landscape to function effectively.
- Long-term predictions include the "main stage" rise of graph learning to address vast quantities of previously ignored relational data, the convergence of hardware and software innovations to reduce model costs, and a shift in AI education toward probability and linear algebra.
- Productivity gains are linked to health improvements through behavioral monitoring, while talent retention is driven by the mission to empower users to exceed perceived capabilities.