Gor Hayrapetyan

Semantic layers promise a single source of truth for metrics, but actually building one reveals a tangle of technical and organizational challenges. In this talk, I'll share my ongoing journey constructing a semantic layer using MetricFlow, dbt, and Snowflake in a multi-tenant data warehouse - and the surprising role AI agents played along the way.
I'll cover three phases. First, agents as builders: how I used AI agents equipped with custom validation tools to accelerate the build process. These agents helped catch metric definition conflicts, validate grain mismatches like fan-out joins, and surface hidden assumptions about schema design. I'll show how running parallel agents with the right tooling turned data migration and validation from a bottleneck into a manageable workflow.
Second, agents as consumers: once a semantic layer exists, it becomes a governed API that agents can query with confidence. I'll demonstrate how a well-structured semantic layer lets agents understand business context - asking "what was revenue last quarter?" and getting trustworthy answers instead of hallucinations.
Third, the honest part: open questions I'm still grappling with. Who should own metric definitions - data engineers, analysts, or domain experts? How do you model the underlying schema to support both human and agent consumers? Which tools actually fit your use case? These aren't solved problems, and I'll share where I am in figuring them out.
This talk is for data engineers, analytics engineers, and anyone curious about how semantic layers and AI agents intersect. You'll leave with practical MetricFlow patterns, lessons from real validation tooling, and a realistic picture of what it takes to build governed data infrastructure with AI as a collaborator.