Data & semantic layer

The metrics track starts from one rule: a metric is defined once, and nothing downstream composes its own SQL. A dbt + MetricFlow semantic layer over a DuckDB warehouse holds nine governed metrics, business rules living in the YAML and mart SQL rather than in any consumer. Two public consumers read that same layer — dbt-semantic-mcp, an MCP server whose three tools let an agent answer KPI questions with lineage back to the dbt nodes, and kpi-console, a SvelteKit application that is itself an MCP client of that server — so the human surface and the agent surface return the same numbers.

The warehouse is a deterministic synthetic set of roughly four thousand orders in one DuckDB file: the patterns transfer, but cloud-warehouse scale and orchestration are out of scope.

The second track is knowledge cartography — a schema-first method that catalogs an unfamiliar domain before any build, plus typed graphs that make a real corpus queryable. The cartography case study carries that method and the graph schemas.