Case study
How a Denver steel fabricator built an AI operating system
Flawless Steel Welding, a commercial structural steel contractor in Denver, runs its daily operations on software we built — estimating, project management, shop production and quality in one system, with an AI agent layer working across all of it. This is what was there before, what was built, and what changed.
The before-state every fabricator knows
The business ran the way most steel shops run: email for decisions, spreadsheets for numbers, detailing and ERP software for their own silos, shared drives for documents, paper on the shop floor, and people remembering the rest. Each tool worked. None of them talked to each other, so the connective tissue of the operation was human effort — re-keying, re-telling, walking over to ask.
That is not a software gap you fix by buying another subscription. Another tool is another silo. The fix was to build the system the business actually runs on, and put AI to work inside it.
The architecture
From tool sprawl to one operating system
Before: six disconnected tool categories — email threads, spreadsheets, detailing and ERP software, shared drives, paper forms, and people remembering things — with nothing connecting them. After: one operating system, where an AI agent layer works across five domains — estimating, project management, production, quality and documents — and the whole system is exposed through MCP to AI assistants like Claude and ChatGPT.
Before — tool sprawl
Nothing talks to anything. Every handoff is a person re-keying, re-telling or re-finding.
After — one operating system
One system of record. Agents work across it; assistants reach it through MCP.
Four workflows, before and after
The results are the operator's own accounting — counts, not marketing percentages. The build facts are verifiable from the public repositories.
Estimating
The problem
Bids lived in spreadsheets — takeoff math in one file, bid status in another, win history in a third, and the numbers re-keyed between them. The current answer was whichever file was open.
What was built
An estimating module: bids, takeoffs and pipeline in one place, with AISC shape math built in instead of pasted in.
The result
Four estimating spreadsheets replaced by one system.
Project management
The problem
Active jobs were tracked in whatever format each project manager preferred — decisions in email threads, documents on a shared drive, the schedule in a spreadsheet. Finding the current state of a job meant asking around.
What was built
A project workspace: every active job with its schedule, documents and decisions in one place, visible to everyone who touches it.
The result
Three separate project trackers consolidated into one workspace.
Production
The problem
Shop floor status lived on paper travelers and in the foreman's head. Knowing what was sitting at which station meant walking the floor or making a call.
What was built
A live work-in-progress board: piece counts by shop station, updated from the floor, visible from the office.
The result
Two daily status meetings replaced by one live WIP board.
Quality
The problem
Quality evidence was paper: inspection records in binders, non-conformance reports on forms, and every audit starting with days of collation.
What was built
A quality module tracking non-conformance reports against AISC 207-25, 360-22 and 303-22, with the evidence attached where the work happened.
The result
One paper NCR binder replaced by tracked non-conformance reports.
The build, in numbers you can check
These are counts from the codebase and release history, not marketing estimates.
10
integrated modules
231
data models
v1.166
shipped releases
Daily
in production use
Screenshots of this system elsewhere on this site are captured from a demonstration instance populated with synthetic data. No client, project, employee or financial information from the operating business appears in them.
Steel AI Assessment — $1,500
Find out what this looks like in your shop
One day inside your operation. You get the AI Opportunity Map — which of your workflows are automatable, and what to build first. The same starting point this case study started from.
Free 30-minute shop or job site walk first • No obligation