I build AI systems for construction teams that protect margin, give teams back capacity, and keep judgment with the people accountable for the work.
More than a decade in commercial construction. Four years building production AI systems. Practical judgment about where automation pays off, where it fails, and where accountable decisions must remain human.
Most AI consultants must learn how contractors think. Most construction consultants cannot build production AI systems. Ten years of experience informs where I look for margin, value, and risk. That is how I build the right systems.
I have run project management and estimating across Division 26 electrical systems, lighting controls, vendor coordination, scope, and pricing, under the bid-day deadlines that make small inconsistencies expensive.
I build and operate the systems myself: agents, retrieval, orchestration, evaluation, local and cloud inference, and the privacy controls required when outputs affect real work.
The portfolio leads with margin protection, validation, and operational capacity. Every case shows where the system stops and accountable judgment begins.
The review reconciles internal quotes, manufacturer quotes, bid breakouts, takeoffs, catalog prices, and service rules. It surfaces disagreements with the evidence attached, then leaves correction and approval to the estimator.
An adversarial QA harness showed that an ambiguous fuzzy matching path was wrong in 63 of 66 cases. The path was removed from automatic pricing, while the validated exact match workflow remained in production.
It turned a 24-line source into an 88-line structured quote, preserving ERP-critical fields, applying catalog pricing from one shared source, and removing the manual steps where prior transcription errors occurred.
Automation earns authority only after it survives contact with real outcomes.
The strain shows up the same way in most contracting businesses:
An estimator doing takeoffs on Saturday because the bid is due Tuesday.
An owner still reviewing drawings at ten at night.
A project manager helping bid work instead of running it.
A vendor quote that lands thirty minutes before the deadline.
The engagement starts with one of those moments, not with a technology.
Stop pricing errors, scope gaps, and inconsistencies before they become change orders, write-offs, or lost revenue. Bid intake, document review, addenda, pricing QA, quote comparison, and estimate-to-operations handoff.
Estimators and project managers handle more work because repetitive coordination, reporting, and document processes run reliably, with explicit exception handling and human review.
Find the highest-value opportunities, define safe operating boundaries, prototype quickly, and walk away from automation that will not pay for itself.
These cases show the production discipline behind the construction work.
A dependency-aware orchestrator produced 2.3x realized end-to-end throughput on decomposable work, with fresh contexts and an independent verify-and-resynthesize gate.
View caseA fail-to-local router and deterministic redaction gate were validated across 1,821 transcripts and 19,124 events with zero confirmed leaks.
View caseA four-stage content pipeline used a separate adversarial reviewer to catch four confidentiality leaks across five drafts before human approval.
View caseA practical sequence for separating true leverage from impressive technology that creates another system to manage.
Map the real workflow, documents, decisions, exceptions, handoffs, and failure points.
Estimate where errors, rework, delay, or administrative load create the greatest commercial cost.
Prototype around one consequential workflow instead of attempting a company-wide transformation.
Measure outcomes, preserve human review, document the process, and expand only when the evidence supports it.
Construction data is not portfolio material. Responsible implementation begins with protecting the business, the people, and the source documents.
Notes on construction systems, AI reliability, operational memory, and how understanding forms inside complex organizations.
A visual essay about retrieval, memory, context, and why collecting more information is not the same as understanding.
The operating principle behind pricing QA, bid review, and every system allowed to influence consequential work.
What cost, independence, contamination, and transaction hardening mean in a production AI system.
Bring one difficult workflow, overloaded process, or recurring document problem. We will determine whether the right answer is process improvement, automation, a focused AI system, or no technology at all.
For initial conversations, describe the workflow without attaching confidential project documents.