CAD AI auditing: traceable window and door inventories from DWG and DXF
A CAD AI deliverable needs more than a sentence saying how many windows a drawing contains. In this project, I worked with controlled DWG input and an official DXF template to produce objects, relationships and evidence that could be traced back to the drawing.
The figures come from my project record, with an evidence date of 13 August 2026. They describe these inputs, not a promise for every drawing. Client drawings were processed in an authorized environment; public illustrations were redacted.
Why an overview screenshot is insufficient
A screenshot-based response reported 7 windows and 5 doors. The object-level workflow reading the original DWG reported 168 window-opening objects and 22 doors. These were different inputs: a PNG does not preserve complete entity handles, coordinates, layers, block membership or proxy payloads. This illustrates information loss, not a fair ranking of models.
If a client asks where the 97th window is, the result needs an object ID and location. A total alone cannot answer that question.
194 candidates versus 190 placed objects
| Item | Recorded result | Meaning |
|---|---|---|
| Candidates | 194 | 4 block-definition templates plus 190 placed objects |
| Placed objects | 190 | 168 windows, 22 doors, none unclassified |
| Location anchors | 190/190 | An anchor for each placed object |
Objects such as opening:38A retain stable handles, classification basis, payload fingerprints and location evidence. A disputed count can be investigated through the object, decoder, basis and derivation record.
The 168 figure refers to window or window-opening objects, not panes of glass. GNU LibreDWG independently confirmed the total of 194 candidates; it did not independently confirm the 168/22 classification. Exact outlines, orientation and bounding boxes remained unfinished. A location anchor is not complete geometry.
A second input: an official DXF template
The other input was the City of San Diego DS-3179 Construction Plan DXF template. My import record contained 187 Benxiang objects and 319 relationships, file version AC1032 and header units of inches. The ten agreed machine constraints passed 10/10.
The record preserved Sheet Index page numbers, discipline codes and titles; references to Greenbook 2024, Whitebook 2024 and California MUTCD 2026; and fields such as NAME OF COMPANY, SITE ADDRESS, PHONE and EMAIL. Recognizing a standard does not establish compliance. The 10/10 result concerns only the agreed checks.
What an auditable result should contain
- Source-file identity, version and SHA-256, identifying the exact input.
- An inventory, classifications, stable IDs and location or drawing evidence.
- Issue lists, totals, machine checks and their sources.
- Uncovered objects and rules, residual risks and conclusions requiring professional review.
Outputs can be agreed as HTML, PDF, CSV, JSON or an .origin package. Table-to-drawing checks, version differences, title blocks and rule prechecks need scope and acceptance criteria before automation is assessed.
Start with one drawing and one question
Tell me what needs checking, what counts as a pass, who will review it and what cannot leave your environment. A sample can establish the inventory, issue examples and evidence before expanding to a drawing set.
A human-workflow blind comparison using the same drawing, scope and criteria has not been completed. I make no tenfold-efficiency claim. Prechecks are not government approval and do not replace qualified professionals' judgment or sign-off.
This approach relates to my Benxiang Protocol article (Chinese) and the 2origin repository. See AI development and delivery for sample assessments.
By Hequbing (He Fangsheng / Dosen). English edition prepared on 2 October 2026, based on my WeChat article (Chinese). Original illustrations and publication records are available there.
Contact: WeChat hecare888; hefangsheng@gmail.com. Services and pricing.