Building Real-Time Context for Data Agents That Improve ML Models

AI agents are changing how machine learning teams work, from exploring data and engineering features to training, evaluating, and debugging models. But agents need more than access to raw data. To be useful and grounded, they need context: feature definitions and lineage, model inputs and outputs, experiment results, and a clear view of how everything connects.

This session explores how to build real-time context for data agents and ML models so agents can move beyond generating ungrounded code to actively help improve models. We'll start with why batch-first infrastructure can slow ML development, covering key challenges including making real-time data available at inference-time, serving at low latency, maintaining consistency between training and serving, and observing feature quality and behavior.

Then, we'll show how to build a shared real-time context layer, powered by a federated query engine that computes and resolves features, and then serves data straight from the source at inference-time in milliseconds. Finally, we'll explore what this changes for data science workflows. With shared context, an agent can investigate changes in model performance, trace predictions back to the data and features behind them, identify missing signals or shifting patterns, and suggest new features, rules, or experiments to test.

The result is agents that can help teams understand what happened, why it happened, and what to try next. The next generation of data science agents will do more than write code faster. With real-time context, they will help teams continuously improve the models behind real-world systems.