Problem
Aggregate accuracy alone does not tell a quality team which defect classes fail most often or which uncertain cases should be reviewed first.
selected case study
Visual QC Project
A steel-defect classification case study using 1,800 NEU-CLS images, handcrafted visual features, model benchmarking, error analysis, and an entropy-ranked review queue, extended with a measured embedding-RAG assistant and a live 3D digital twin of the inspection line.
Quick technical review
The case study in four points.
Architecture
A deterministic 80/20 holdout evaluates HOG, Gabor, and grid features across baseline, Logistic Regression, and Random Forest models.
Results
1,800 NEU-CLS images across 6 classes
Verification path
Start with the case study and model-comparison artifacts before reviewing implementation details.
Summaries reuse the project’s documented evidence. Review the complete case study for evaluation boundaries.
Evidence-led visual review
Explore the actual work.
Actual project screenshots and repository figures are shown with source provenance.
Dated demo screenshot
Public demonstration screenshot dated 2026-10-14; does not establish that the demo is currently online.
Open published source1,800 NEU-CLS images across 6 classes
0.9389 accuracy / 0.9392 macro F1
Top 20% review queue captures 81.8% of errors
Project proof
Verified external evidence
Last verified 2026-10-05
dashboard proof
Live HMI with 3D line twin
dashboard proof
RAG map
dashboard proof
SPC p-chart
dashboard proof
In-browser ResNet-18 demo (Hugging Face Space)
repo proof
Repository
document proof
RAG case study
document proof
Kaggle notebook: 100% accuracy and where it ends
document proof
ResNet-18 model card on Hugging Face
Loading the interactive view...
Problem and approach
Problem
Aggregate accuracy alone does not tell a quality team which defect classes fail most often or which uncertain cases should be reviewed first.
Architecture
A deterministic 80/20 holdout evaluates HOG, Gabor, and grid features across baseline, Logistic Regression, and Random Forest models.
Review logic
Prediction entropy ranks low-confidence cases so a limited human-review budget captures a disproportionate share of model errors.
AI engineering layer
BGE-base embeddings served from Chroma were chosen from a four-way retriever benchmark; a local Qwen2.5 generator and prompt were selected by NLI-measured faithfulness against a no-retrieval control.
Why it matters
It connects computer-vision evaluation, class-level error analysis, and a practical human-in-the-loop quality-review strategy.
Architecture and analysis
- Evaluates 1,800 balanced NEU-CLS images across six steel-surface defect classes.
- Uses Pareto and class-level analysis to identify where observed errors concentrate.
- Turns model uncertainty into a review queue instead of presenting predictions without an operating strategy.
- Shows RAG retrieval on a 3D embedding map whose highlights come from the real search, with the projection distortion reported on the page.
- Renders the simulated line as a live 3D digital twin: scan gantry, reject pusher with NEU defect classes, OK stacking.
- Lets an operator confirm or correct the predicted defect of each rejected unit, with decisions persisted per visitor and exported as a quality-engineer CSV.
- Adds an SPC p-chart of the reject rate per batch with 3-sigma control limits, and a CNN pipeline (ResNet18 / EfficientNet-B0) built to be compared with the Random Forest on the exact same holdout.
Testing and validation
- Operator review queue with Postgres (SQLite fallback) persistence, versioned migrations, row quotas and per-client rate limits on the public write endpoints
- CNN comparison pipeline with temperature scaling, ECE, bootstrap intervals, Grad-CAM and ONNX export, verified on synthetic data; a separate fine-tuned ResNet-18 baseline (cnn/train_cnn.py) scores 1.000 accuracy on the random and on a blocked 80/20 split of NEU-CLS, which the repository reads as a saturated benchmark rather than production readiness
- Dockerfile and docker-compose (app + Postgres); CI with ruff, mypy, an 80% coverage floor, a Postgres service and Playwright end-to-end tests on Python 3.12
- HOG, Gabor filter-bank, and grid-statistics feature pipeline
- Model benchmark, confusion matrix, class metrics, and Pareto error artifacts
- Entropy-ranked review-queue analysis with versioned JSON evidence
- RAG pipeline: pinned Wikipedia corpus, retriever benchmark, Chroma index, NLI faithfulness, UMAP projection
- Three.js line twin synced to the simulation API, with a 2D fallback
Operational value
- Achieves 0.9389 accuracy and 0.9392 macro F1 with the selected Random Forest model.
- Shows that two defect classes account for approximately 64% of observed errors.
- The top 20% entropy-ranked review queue captures 81.8% of model errors.
- RAG retrieval puts the exact source passage in the top 4 for 94.8% of 77 questions (TF-IDF baseline: 89.6%).
- 75.7% of generated answer sentences are entailed by the retrieved passages, versus 20.7% for the same model without retrieval.
- A fine-tuned ImageNet ResNet-18 reaches 1.000 accuracy (360/360) on both the random and the blocked holdout, so NEU-CLS no longer separates models; under test-time degradation its accuracy falls to roughly 0.34-0.39 with strong sensor noise and 0.36-0.38 with strong blur.
How to inspect
- Start with the case study and model-comparison artifacts before reviewing implementation details.
- Inspect the confusion matrix and Pareto analysis to understand class-level failure modes.
- Use the entropy review-queue curve as the main operational decision surface.
- Rotate the embedding map on this page and pick a sample question: the highlighted passages are the real 768-d top 4, not the nearest dots on screen.