▶  WATCH THE FULL PRESENTATION

Computer Vision Is the Andon Cord Construction Has Always Needed
LCI Conference 2025  ·  Reid Senescu, Doxel  &  Mike Miller, DPR Construction

A Disagreement That Turned Into a Discovery

Doxel's system was designed to track progress, but on a hyperscale data center project with DPR Construction, it caught something that no daily report, RFI, or schedule update had flagged, and the lesson that came out of it changed how the team interpreted data discrepancies entirely.

The story starts with a flag. Doxel's AI detected uninstalled security components near certain doors. The electrical trade partner pushed back hard, claiming they had roughed in all the security to those doors. In their estimation, the work was done. 

After further investigation, the team found the truth: the security boxes had been installed. Three feet to the right of where they were supposed to be.

Wrong Location. Not Flagged Anywhere.

The components had been physically installed, but they were mislocated relative to the BIM. When comparing the 360° site photos against the model, Doxel’s AI correctly identified them as not installed in the designated location.

"If something's showing as not installed and the trade partner says it's installed, we probably have a quality control problem. Not the intended use case — but awesome to see."
‍— Mike Miller, Superintendent, DPR Construction

The team had stumbled onto a new interpretive principle. When Doxel flags something as missing and the trade says it's done, don't default to assuming the data is wrong. Investigate. The discrepancy might not be a tracking error; it might be a quality flag.

The Cost of Dismissing the Signal

Mike was direct about what happened next and what it cost. Rework followed.  But by investigating when they did, the team headed off even higher costs than if the issue had been found later. 

REWORK WARNING:  Dismissing data because it contradicts expectation is how quality issues get buried. The cost of investigation is almost always lower than the cost of rework — especially once walls are closed.

This is not an abstract lean principle. It played out on a real job, on real infrastructure, with real rework costs. The lesson is practical: when scan data and field reports disagree, treat the disagreement as information, not noise.

What This Means for Quality Management on Complex Projects

Construction quality management has traditionally relied on scheduled inspections, trade self-reporting, and periodic walkthroughs. These methods work reasonably well for obvious defects. They are poor at catching components that are physically present, but are installed in the wrong place relative to the design.

Computer vision can fill this gap by comparing what is physically present against the BIM at the component level across all visible trades every week. Mislocations look identical to missing components from the system's perspective, because in both cases, the component is not where it should be.

The practical recommendation from Mike's experience is to establish a protocol for investigating discrepancies rather than defaulting to dismissal. When a trade reports complete and the system reports incomplete, send someone to review the discrepancy. It only takes minutes, but it can prevent weeks of rework.

Beyond Rework: The Documentation Value

There is a secondary benefit this story highlights: objective, time-stamped documentation of installation location for every component. On a complex facility like a data center, where systems are dense, and modifications may be needed years later, having a record of where things were actually installed, not just where they were designed to go, has ongoing operational value.

This use case wasn't in the sales deck. It emerged from a real disagreement on a real job. That's often how the most durable capabilities get discovered.