Matthew Mullins

This session makes the case that the semantic layer is the control point for agentic data work, and looks at three benefits that show up in production. Context: grounding agents in defined entities, metrics, and relationships so they generate correct queries instead of plausible ones. Governance: enforcing policy, lineage, and access at the semantic boundary rather than hoping the prompt holds. Cost: cutting token spend and warehouse burn by giving agents a compact, reusable map instead of letting them explore raw schemas by trial and error. Attendees will leave with a practical view of what to put in a semantic layer before pointing agents at their data, and how to tell a genuine one from a rebranded metrics catalogue.

DataOps is often framed as a collection of tools. In practice it is a culture and a set of engineering behaviors adapted from software and platform teams to the realities of data work.

We propose and demonstrate an architecture that swaps compute engines by workload size. Using Apache Iceberg™ as a central store, end users can query small-to-moderate data locally with DuckDB in the browser but automatically switch to Snowflake for large-scale queries. We address predicate and projection pushdown, local caching and SQL transpilling. This mixed-engine approach is flexible, reduces cloud costs and scales effortlessly.