From Weeks to Hours: Engineering an AI Pipeline for Site Investigation Reports
The whole series on one page, from the business problem to the lakehouse roadmap.
Categories
Everything on this blog falls into four buckets, one per stage of the two series
Azure AI & Agents
From Weeks to Hours: Why a Site Investigation Report Is Not a Chatbot Problem
The business problem, why an LLM alone is the wrong answer, and the end-to-end pipeline.
The Model Narrates, Code Decides
Provenance, reason_null, multiplicity guards, divergences, and the grounding gate.
Hosted Agents on Azure AI Foundry
Per-agent identities, the attach-only Site object, two human pauses, one trace, one Bicep template.
The Data Layer: Authoritative Sources Wrapped as Guarded Tools
Regrid, FEMA, USDA, wetlands, Planetary Computer, and a jurisdiction-scoped ordinance corpus.
Trust Engineering: The Guardrails Are the Product
Offline tests, runtime gates, production evaluation, business impact, lessons, and the roadmap.
Fabric & AI
From Raw Data to Stunning Report: The Complete Series
All five parts on one page, from OneLake bronze to a Figma-themed Power BI report.
Load Data into the Lakehouse with Copilot + Fabric MCP
Shortcuts, a SQL snapshot second source, and Fabric MCP setup with your own login.
Data Engineering: Bronze to Silver to Gold
Joining shortcut CSVs with the SQL snapshot, quarantine, and gold star schema.
Power BI & AI
Semantic Models: Why They Matter + Build with Copilot
DirectLake, DAX measures, and RLS via Power BI MCP.
Build the Report Locally with Copilot
PBIP in VS Code, Desktop Bridge reload and screenshot loop, GitHub Actions deploy.
Design & AI
Make It Beautiful: Figma Design System to Stunning Report
Figma MCP tokens to Power BI theme, layout polish, and the production checklist.