InspectLoop

Scratch fixture — OpenCV findings and HITL confirm
Foreign-object fixture — viewpoint and stop-line confirm
Mis-assembly fixture — rotated part, stop-line confirm

InspectLoop is an industrial QA inspection agent for the OpenCV AI Competition 2026 Agentic Vision track.

Camera frames run through an OpenCV 5 CPU pipeline that detects scratches, foreign objects, and mis-assembly against a golden template. Findings drive MCP-style tools (zoom, change_viewpoint, open_work_order, stop_line) with human-in-the-loop Approve/Reject before any stop-line action, plus an observable perceive→decide→confirm→act trace.

Live demo (no login): https://inspectloop.vercel.app/demo — pick a synthetic fixture or upload a frame; Approve/Reject updates tickets and the line banner from the API. Health: OpenCV 5.0.0.

OpenCV 5 usage: thresholding, morphology, findContours, minAreaRect, absdiff, matchTemplate, annotation (opencv-python 5.0.0.93). Eval on synthetic fixtures: 4/4 (ok / scratch / foreign / misassembly).

AWS: same FastAPI image is packaged for App Runner + optional S3 media (see aws/ in the repo). Current public demo runs on Vercel; App Runner HOWTO ships for free-tier deploy.

Explainer (EN, ~2.7 min): https://review-clips.vercel.app/v/opencv-inspectloop-explainer-en Short pitch (EN, 18s): https://review-clips.vercel.app/v/opencv-inspectloop-pitch-en

Built by Caesar (er-s-an). Synthetic fixtures only — do not wire to a physical E-stop without safety engineering.