I go where the
process is broken.

The process that wouldn't run

A large fiber-infrastructure customer needed to catch double-claimed footage in engineering design data. Each design line claims a stretch of a construction route — feet 100 to 250, say. When two lines claim the same stretch, that footage gets counted twice, and every cost and quantity rollup downstream is wrong. The check existed only as a Power BI process that had stopped completing, so it fell back to people reconciling by hand, four to eight hours at a time, every import.

I rebuilt it as a native pipeline: indexed lookups instead of full-table scans, and a chunker that reads the remaining query budget and sizes itself to fit inside it. When a stretch of route comes back dense enough to threaten the row cap, it halves its own chunk width and picks up from the last row it saw. A separate guard watches heap against a soft ceiling set below the platform's hard limit. When that trips, the affected chunk fails with a message telling the operator what to narrow — instead of taking the whole job down at the hard limit.

Four batch phases folded into one configurable job, and tuning moved into metadata, so admins change thresholds and modes without a deployment.

Two terabytes of spreadsheet

An entire contractor payment operation lived in Excel — engineering drawings, GIS data, and field work logs, cross-checked by hand before anyone could be paid. Roughly two terabytes of it. Every error was a payment error, and every audit was an archaeology project.

I architected the platform that replaced it and led the two-engineer core build: ingest the engineering records alongside contractor-submitted logs, validate the work actually done against what was specified, and post reconciled records straight through to finance. The audit trail stopped being something reconstructed after the fact.

Everyone used AI. Nobody the same way.

Across the engineering org, AI assistance had gone from novelty to daily habit without anyone agreeing on how. Output quality tracked whoever happened to write the prompt. The gap wasn't enthusiasm — it was that the models had no reliable access to how we actually build.

I built an internal Model Context Protocol server and an agent rule set encoding our conventions. Turning that into something the whole org actually used was a team effort — engineers across teams stress-tested the rules, pushed back on what didn't hold up in real work, and shaped what the tooling should enforce. The interesting part was never the wiring. It was the collective judgment about what models should not be trusted with, and where a human review gate genuinely earns its cost.