Open workbench
This page separates measured progress from ideas still under investigation.
A test schedule contains 17 door rows after its headers and footer are excluded. Image-guided extraction previously returned 15 unique records in a row-by-row run. The current batched experiment attaches source-row-band identifiers so a missing record can be traced to its specific image band. This is promising, not a benchmark claim: field-level verification and repeatable evaluation remain required.
Testing began with sheet A8.2 and focuses on door rows, dimensions, materials, hardware groups, frame types, ratings, and remarks. Merged cells and hardware-group blocks remain important stress cases.
Table detection, row count, field extraction, normalized values, and final reviewed records are separate measures. The next benchmark should report each one so improvements remain meaningful.
Models will change. A collection of hand-verified construction schedules, paired with clear schemas and source evidence, makes future models easier to compare and replace.
Roadmap
Small, measurable steps before broad automation.
Document the exact model, prompt, settings, output, review, and score.
Connect every extracted field to its sheet region and review state.
Use the drawing index to route likely schedule sheets to the right schema.
Compare schedules, specifications, BIM records, and drawing revisions.
Planned working example
The Test Lab lets authenticated Hugging Face users try one focused version of the pipeline themselves. Shared ZeroGPU infrastructure supplies the temporary GPU; the user only needs a browser and a safe test image. Capacity is shared, so a GPU queue or time limit can occur during active testing.
Use a Hugging Face account so RealEyesVR does not store passwords.
Submit one public, synthetic, or otherwise authorized schedule image.
ZeroGPU temporarily allocates shared GPU capacity for extraction.
Review structured rows, source-row-band warnings, evidence, and raw JSON.
One image, one known schema, and one measurable result. Broader drawing types come only after this workflow proves repeatable.
Processing time, row count, source-row coverage, field accuracy, validation warnings, corrections, and whether the result was useful without expert guidance.
The Test Lab will use cloud processing. Visitors must not upload client, employer, personal, or otherwise restricted information.
Follow the prototype
The first release will favor a small, understandable experiment over a broad claim of construction-document automation.