A held-out question set is a real release control for analytics agents. It is not proof the agent is safe. What the score actually measures, and what to require …
Blog
Guides, deep-dives, and experiments in semantic modeling, data engineering, and AI.
Agents are easy to demo. Trustworthy meaning, access, checks, lineage, and ownership are still scarce. That boring layer is the real bet.
Why data teams must move beyond prompt-driven AI to spec-driven, lineage-validated deployments—and how dbt, contracts, and metadata make that possible.
SQL is not the bottleneck for most teams; understanding the data is. Why profiling and governed context must come before AI-generated models.
Semantic layers are becoming the controlled, foundational “business memory” powering reliable AI decisions.
Semantic layers and knowledge graphs are not rivals: one governs metrics and repeatable numbers; the other structures entities and relationships. AI needs both …