Explainable Visual Defect Inspection Agent for Small Manufacturing

Audience

Who it is for

Teams comfortable with one vision model plus a small agent wrapper — hardware is a webcam only.

Webcam or upload demo, labelled class, heatmap overlay, PDF of the last 10 inspections.

Branches

Team size: 1–4 students

Problem

Existing vs proposed

Small units still do 100% visual QC. Misses are costly; a black-box model is not acceptable on a factory floor or in viva.

CNN/transfer-learning classifier, Grad-CAM-style map, LLM that may only describe the highlighted region, lot log, and a “needs human” low-confidence path.

Build

Modules

1. 1. Data

   • Public defect set + train/val split
   • Augmentation, leakage check

2. 2. Model

   • Transfer learning
   • Calibration / confidence
   • Heatmap

3. 3. Agent note

   • Describe only the highlighted region
   • Refuse if confidence is low

4. 4. QC desk

   • Lot ID, operator, pass/fail override
   • Shift report PDF

Deliverables

Included with the kit

• Complete source code (ZIP + GitHub-style repo)
• Database schema and sample data / dataset notes
• README.md — how to run, stack, modules, and demo path (Windows + college lab)
• Viva question bank with short answers
• Two mentor reviews: architecture, then final demo