Research
The questions the lab is organized around, with why each one matters and what the lab plans to test.
Open questions
Private document AI for RFIs and submittals
Can AI read, check, and route RFIs, submittals, and change orders accurately while running entirely on hardware the operator owns?
Why it matters
Document-heavy workflows are a common bottleneck in project execution, and many construction organizations will not send operational data to public AI services. If document AI only works in someone else’s cloud, these teams cannot use it.
What we plan to test
Whether models running on-premise, or in a private cloud the operator controls, can extract, check, and route these documents well enough that a person reviews only the items that need judgment.
Field data capture by voice and photo
Can crews capture daily logs, site photos, and safety checks by voice and photo on the devices they already carry, without learning a new app?
Why it matters
Field information is often entered late, inconsistently, or not at all. New interfaces meet resistance on site, so tools that ask crews to change how they work tend to go unused.
What we plan to test
Whether capture through channels crews already use, such as voice notes, text messages, and photos, produces structured records the office can rely on.
Early risk signals from routine project data
Can the records projects already produce, such as daily logs, RFIs, change orders, and meeting minutes, show schedule and cost risk early enough to act on?
Why it matters
Leadership often sees project status through reports compiled by hand, which are out of date by the time they arrive. Risks are hard to compare across projects.
What we plan to test
Whether signals drawn from routine records, processed inside the operator’s environment, flag emerging risk earlier than hand-compiled reports do.
Evaluating AI in safety-critical construction workflows
How should AI be evaluated before it is trusted with safety forms, inspections, and other workflows where an error has physical consequences?
Why it matters
The lab uses automation only where it gives a clear, measurable advantage. In safety-critical work a wrong answer costs more than in back-office work, so general accuracy claims are not enough.
What we plan to test
Evaluation methods that measure errors on real workflow records, keep a person in the loop, and leave an audit trail that can be inspected afterwards.
Grounded in field work
Field problems from real construction engagements shaped this agenda. That commercial work is now done by CASA Media, and client details stay private.
Open methods, public findings
- Open methods. Each study will describe how it is designed and evaluated, so others can check and repeat it.
- Public findings. The lab will publish what it learns, including results that do not work out.
- Operator-owned infrastructure. The systems we study will run on hardware the operator owns or in a private cloud the operator controls, and operational data will stay there.
Have a field problem that fits one of these questions? Tell us about it.
Partner with the lab