In a single-tenant shop, a clumsy semantic model is an annoyance. In a SaaS ISV with thousands of customers on shared Fabric capacity, it's a defect replicated thousands of times — opaque field names degrading Copilot, missing lineage breaking trust, a silent DAX bug shipping wrong numbers to every tenant, one runaway Spark job throttling them all. At ISV scale, manual data engineering doesn't add up — it multiplies.

This is the data-platform architecture that breaks that curve. We treat Fabric artifacts as source code — TMDL/TMSL/PBIR (the same tabular semantics as SQL Server Analysis Services) round-tripped through deterministic, testable cores — and automate the lifecycle end to end: Discover → Author → Audit → Modify → Publish → Govern, across a Lakehouse/Warehouse estate fed by Dataflows Gen2 and Data Factory.

Two pillars carry it. MCP: one server, ~35 typed tools mirroring a single service layer 1:1, so agents and humans share one capability set. Skills: versioned, governed capability packs (modeling, DAX, accessibility/BPA audits, domain-grounded metadata) that encode expert practice as reusable units — your best practices in tooling, not in three people's heads.

Then the parts that matter at scale: lineage and cataloging beyond basic Purview, data-quality gates, capacity governance, multi-tenant isolation, secret less identity, and "deterministic by default, agentic by preference" so a write to a customer's model is always safe.