Advisory, dev studio, and field notes on Fabric, Power BI, Copilot and Azure AI agents. I help teams ship data work that holds up in production, and I write down what breaks along the way.
The whole story on one page: why a liability document can't be a chatbot, hosted agents on Azure AI Foundry, every fact on a cited deterministic path, and an engineer holding the pen. Or read it as the five-part series below…
What a Site Investigation Report is, why it takes an engineer weeks, why "point an LLM at it" is wrong, and the end-to-end pipeline that replaces the copy-paste.
The boundary the whole design defends: provenance on every value, a reason on every gap, no collapsed parcels, conflicts as divergences, and a machine-checked summary.
Per-agent Entra identities, an attach-only Site object every hop must pass, two deliberate human pauses, and provenance that doubles as the OpenTelemetry trace.
Regrid, FEMA, USDA, wetlands and Planetary Computer behind guarded tools, plus an ordinance corpus in Azure AI Search that can only cite the parcel's own jurisdiction.
The finale: offline tests, runtime gates that halt instead of degrade, production evaluation, the business economics, six lessons, and where the roadmap goes next.
Start where every good report starts: raw files to OneLake bronze. Set up Fabric MCP, create shortcuts, and let Copilot build your ingestion pipeline.
Copilot-written Spark notebooks clean bronze into silver, quarantine bad rows, and build gold star-schema tables ready for modeling.
Why models beat raw tables: one definition of sales, DirectLake speed, DAX measures + RLS, all generated via Power BI MCP.
PBIP in VS Code on your Part 3 model: Copilot generates pages, binds visuals, wires interactions. Functional first, beauty next.
The payoff: Figma MCP pulls your design tokens, Copilot themes the report, layout matches Figma frames. Ship the beautiful outcome.