ProofShape
Phone-based coarse 3D inspection for manufactured parts, comparing a supplier's photographed part against the buyer's CAD file.
Aug 2026
Overview
CS595 capstone, started August 2026, three-person team.
A supplier photographs a part on a printed reference sheet, guided by on-screen instructions. The system reconstructs a metric 3D model from the photos and compares it against the buyer's CAD file, returning a Pass / Fail / Rescan / Unverifiable report with a deviation heatmap.
The report also includes a provenance mask that separates the surface actually measured from the camera from surface that was AI-completed to fill gaps the photos didn't cover — a Stereon-adjacent idea (see Stereon) applied here to distinguishing observed geometry from inferred geometry, not to authenticity detection.
Milestones
- Authored and froze a versioned OpenAPI 3.1 contract decoupling the capture and reconstruction services, with per-endpoint fixtures and schema-validation tests — including a structural regression test enforcing the system's core safety invariant, that unobserved regions can never contribute to a Pass verdict.
- Collected and validated the golden-capture benchmark dataset: five multi-view sets (~140 phone photos) of rigid parts on a ChArUco board across lighting conditions, with caliper-derived CAD ground truth.
- Established the team's engineering baseline — enforced branch protection, required CI (lint and type checks), one-PR-per-story review workflow — and reviewed teammates' pull requests, catching data-integrity and cross-platform defects before merge.
Status
In progress, started August 2026.
Stack
Python, PyTorch, Next.js, TypeScript, OpenAPI 3.1, 3D reconstruction, generative completion, OpenCV