Construction data twin research
Construction drawings are full of data. Let’s make it usable.
RealEyesVR is an independent project exploring how visual AI can read drawing schedules, check what it found, and turn the results into structured data people can actually query.
Built from a VDC/BIM perspective. Tested on real construction workflows.
The problem
The information exists. It is simply trapped.
Door schedules, equipment lists, panelboards, finishes, and hardware groups contain information teams repeatedly re-enter by hand. PDFs are easy to view, but difficult to query, compare, or connect to the rest of a project.
Today
People search sheets, read tiny tables, copy values, and rebuild the same lists in spreadsheets and models.
The idea
Let visual AI perform the first reading, then validate every record and preserve a direct link back to its source.
The result
A living data layer that can support estimating, procurement, coordination, field checks, and future analysis.
The workflow
From sheet to trusted record
AI starts the process. It does not get the final word.
- 01
Read
Qwen 3.5 examines the drawing and identifies useful schedules and fields.
- 02
Structure
Results are converted into a consistent schema with normalized names and units.
- 03
Check
Rules catch missing values, impossible results, duplicates, and broken references.
- 04
Review
Ambiguous records receive a second AI review or a focused human check.
- 05
Connect
Verified records enter the data twin with their sheet, region, and revision attached.
The destination
Not another PDF viewer. A project data layer.
Once drawing information is structured, teams can ask practical questions:
- How many hollow-metal doors use hardware group HW-1?
- Which scheduled items changed between drawing revisions?
- Which equipment records are missing electrical requirements?
- Where do drawing schedules and the BIM model disagree?
Current status
A working research project—not a finished product.
Recent tests show that Qwen 3.5 can serve as an effective first visual gate, followed by deterministic checks and targeted review. The next work is making that success repeatable across projects, disciplines, and revisions.
- Door and frame schedules
- Hardware groups
- Structured JSON output
- Source-linked validation
- ZeroGPU community Test Lab