Blog
Notes on governed metrics
How we model, query and secure the numbers that dashboards and AI agents share. From the people building datahashi.
Governance & trustThe engine never opens your warehouse
Why datahashi stopped holding warehouse passwords — compile and execute are different jobs, and agents make the wrong trust boundary expensive.
Agents & MCPWhy LLMs invent joins (and how to stop them)
Explain why models invent joins from schemas and how governed metrics remove join invention from the agent loop.
QueryingOne request shape for every consumer
Dashboards, APIs, and agents should ask for metrics the same way. One structured request shape keeps definitions stable and tooling simple.
Strategy & comparisonsSemantic layer vs dbt models vs BI cubes
Compare dbt models, BI cubes, and a serving semantic layer so you know what each owns and why agents need a request-shaped contract on top.
Agents & MCPWhat is a semantic layer (and why agents need one)
A semantic layer gives agents stable metric definitions, governed access, and one query shape so AI systems stop inventing their own numbers.
Strategy & comparisonsFrom BI-only semantic layer to agent-ready serving
How to move from a BI-only semantic layer to agent-ready serving without rewriting every dashboard.
QueryingQuerying the semantic layer via REST
Call the semantic layer over REST with a metrics × dimensions × filters payload and read certified results - conceptual request and response shapes.
Agents & MCPMulti-tool MCP chaos: one revenue definition
A composite war story about conflicting revenue answers across MCP tools and how one certified metric restored trust.
Warehouses & setupPostgres or MySQL as a starter warehouse
Use Postgres or MySQL as a starter warehouse for a thin semantic layer, then move to Snowflake or BigQuery without rewriting consumers.
Modeling & metricsModeling revenue so agents cannot break it
How to define revenue as a certified metric with clear grain, filters, and guardrails so agents and humans get the same number.
Strategy & comparisonsWhen you do not need a semantic layer yet
Honest signals for postponing a semantic layer versus when agent and multi-consumer chaos makes one urgent.
Governance & trust