Data foundation & agentic evaluation

This is the practical side of the data leg of my practice. Every team I have worked with has the same problem: the numbers exist, but they live in five tools, nobody trusts them, and the person who needs the answer cannot write the query.

How it is built

• Foundation: connectors pull raw events, content and revenue data into a modeled warehouse layer.
• Semantics: a typed metric layer so "revenue" means one thing, everywhere.
• Access: natural language questions resolve against the semantic layer, never against raw tables.
• Agents: scheduled and triggered runs propose or take actions on top of the same layer.
• Evaluation: every run is scored against recorded outcomes, so quality is measurable.

The interesting part is not the chat. It is that the chat is only trustworthy because the layer underneath it is boring, typed and tested.

It is early, and deliberately unnamed. I am building it in the open — the reasoning lives on the Data page.