WEXSPACE AI — Agentic Room Vision with OpenCV 5

Inspiration

Engineering students, designers, and small contractors often need a transparent first estimate from a room image, but an image alone does not establish a trustworthy physical scale. WEXSPACE AI explores an evidence-gated workflow that connects computer vision to a real quantity-calculation tool while keeping the human responsible for validating dimensions.

What it does

The current technical prototype detects a synthetic, scale-calibrated rectangular room outline with OpenCV 5, proposes provisional length and width, and routes the result through a human-confirmation gate before an MCP tool may update the measured-room project. If scale is absent, perspective is uncertain, the image cannot be decoded, or a room requirement conflicts, it stops the tool call and requests the appropriate correction.

For the controlled 6.04 m × 4.04 m fixture, the provider-backed calculation returned 24.4016 m² floor area, 55.8 m² net wall area, 12.276 L paint, and 75 floor tiles, under explicit user-supplied assumptions. These are preliminary estimates, not certified measurements or purchases.

How it works

Perception: OpenCV 5.0.0 in a Python environment on an existing AWS EC2 instance processes an image with a known synthetic metric scale. Decision: A Python agent bridge chooses a tool call or requests scale, rectification, valid image, or human confirmation. Action: After fixture confirmation, MCP SDK 2.3.0 invokes the actual Node.js measured-room engine through its exposed tools (inspect_project, update_room, export_dossier). Readback: An independent HTTP read verifies the resulting revision and quantities. Replaying the same operation returns a cached receipt without a second update.

Technical verification

A scoped AWS execution receipt records OpenCV 5.0.0 and four passing vision tests, MCP protocol negotiation (2025-11-25), one real backend revision 0 → 1, and readback of the calculated quantities. The evidence uses synthetic images and assumptions; it does not prove real-room metrology, hosted visual UI integration, a live ChatGPT image invocation, or a completed public demonstration.

Challenges and responsible use

Pixel geometry is not a reliable real-world measurement without verified scale and perspective. The prototype deliberately blocks unsupported estimates and requires confirmation before state changes. No construction work, procurement, or payments are triggered.

Next development

Integrate the visual workflow into the hosted room interface, test with independently measured real images, document calibration error and failure rates, and record a judge-accessible live demo. These items are not yet claimed complete.

Judge-accessible source and reproducibility

Public source branch: https://github.com/moatazfarea/wexspace-engineering-design-workflow-studio/tree/feature/opencv5-agentic-room-20261026-r01

For the measured-room web application: use Node.js 22+, run node server.mjs, open http://127.0.0.1:4188/, and run node --test tests/*.test.mjs. For the OpenCV 5 agentic proof, see opencv_candidate/README.md, opencv_candidate/AWS_QUALIFICATION.md, and the bounded AWS receipts under evidence/. The synthetic image pipeline and the hosted browser UI are separate demonstrations; the hosted interface does not yet accept an image upload as a live computer-vision input. The OpenCV runtime was independently read back on the existing AWS EC2 instance; the repo provides source and evidence, not a claim of production-grade real-room measurement.

This source branch is public and judge-readable. A contest video and the final submission are still pending; publishing this project page is not the same as submitting it to the competition.

Built With

• aws-ec2
• javascript
• mcp
• node.js
• opencv-5
• python