Human-guided AI · Reverse-engineered 3D
From mixed evidence.
To a verified 3D asset.
A disciplined reconstruction workflow turned product photos, assembly drawings, and human corrections into explicit geometry, transparent QA, a frozen GLB, and an interactive browser delivery.
Independent visual reconstruction. Public depiction and model-distribution rights are approved; source references remain unpublished.
01 · Human-in-the-loop method
The model improved because evidence and judgment stayed separate.
The human defined what must look right and corrected ambiguous visual readings. The AI translated those decisions into task constraints, scene-spec fields, procedural geometry, repeatable cameras, and validation reports. Neither role was treated as optional.
Collect evidence
Separate product views, assembly diagrams, instruction pages, and low-confidence unseen regions.
Write the contract
Store dimensions, units, sources, confidence, and assumptions in a parseable scene specification.
Build and align
Generate the blockout, compare orthographic views, mark offsets, then make bounded geometry corrections.
Freeze decisions
Snapshot approved geometry and later visual assets with reports, SHA-256 hashes, and read-only archives.
Validate delivery
Round-trip the GLB, verify UV and material structure, run browser QA, and state limitations instead of hiding them.
02 · Alignment before detail
Orthographic evidence made disagreements visible before the model changed.
Front and driveway-side passes used fixed cameras and visual review notes to separate position, size, curvature, and opening-layout errors. The most important failure was semantic: the roof billboard and curved storefront awning were initially confused. Human clarification corrected the category before more geometry was built.
03 · Freeze as a production primitive
Approval became a file set, not a memory.
The final milestone records the approved GLB, scene-spec version, six QA views, material inventory, pipeline report, rights boundary, limitations, and hashes. This creates a stable base for future work and makes silent regressions detectable.
ASSETVisual Asset v1.0
VALIDATIONGLB round-trip · 0 invalid transforms
UV216 / 216 meshes valid
BOUNDARYPrivate evidence excluded from public case
04 · Material architecture
Recognition details were organized without turning the asset into a texture dump.
Ten explicit slots keep walls, roof, brick, metal, glass, ground, signage, props, trunk, and foliage independently editable. UV checks and a separate signage region preserve direction and future revision safety.
05 · Interactive delivery
Inspect the frozen public model in real time.
Drag to orbit and switch between the same perspective and orthographic camera families used for QA. The public GLB keeps the approved geometry, UVs, materials, and textures while removing private source metadata.
The download contains the complete frozen geometry, UVs, materials, and embedded textures; workstation metadata is removed.
06 · Technical evidence
Visual approval and technical validation were recorded as different gates.
The asset passed production build, browser QA, GLB round-trip, material and UV checks, duplicate-name checks, transform checks, bounds checks, and final freeze verification. Product fidelity remains a human visual decision.
Technical milestone: passed · Public depiction and model distribution: approved
07 · Limitations and rights
A transparent visual reconstruction is still not an official or survey-grade asset.
Conservative rear facade
Rear evidence was limited, so the final facade intentionally remains a lower-confidence conservative reconstruction.
Visual milestone scope
No interior, animation, LOD, collider, emissive, normal-map, grime, or manufacturing-precision pass is included.
Public depiction rights
Public depiction and distribution of the reconstructed model are approved. Product photos and instruction materials remain source evidence and are not republished.
Independent reconstruction
The approved public model remains an independently produced reconstruction and is not presented as an official manufacturer asset.
Workflow result
AI accelerated construction. Human review made the asset trustworthy.
The complete source, reports, reference evidence, and immutable freeze remain in a private repository. This public case shares the process, generated QA evidence, limitations, and the downloadable approved model with workstation metadata removed.