Business Context
VoltLens manufactures consumer electronics and uses a vision model to detect visible assembly defects from product-line images. The current model misses too many defects and is too slow for reliable deployment on the inspection line.
Dataset
You are given a labeled image classification dataset for a binary computer vision task: defective vs non-defective product images.
| Feature Group | Count | Examples |
|---|
| RGB images | 1 input tensor | 224x224x3 product photos from 6 factory cameras |
| Metadata | 4 | camera_id, production_line, shift, product_type |
| Labels | 1 target | defect_present (0/1) |
- Size: 120K images collected over 9 months
- Target: Binary classification — defect present (1) vs no visible defect (0)
- Class balance: 8% defective, 92% non-defective
- Missing data: ~3% missing metadata fields; image quality varies due to blur, glare, and lighting shifts
- Data issues: Near-duplicate frames from the same product and temporal drift after camera recalibration
Success Criteria
A production-ready solution should achieve strong recall on defective items while keeping false alarms manageable. Good enough means:
- Recall on defects >= 0.92
- Precision >= 0.70 at the selected operating threshold
- PR-AUC >= 0.85 on a held-out test set
- P95 inference latency < 50 ms/image on a single T4 GPU or equivalent
Constraints
- The model will be used in near-real-time on the factory line
- False negatives are more costly than false positives
- The quality team needs image-level explanations for flagged defects
- Retraining budget is limited to a weekly batch job
Deliverables
- Build and optimize a computer vision model for defect detection
- Define a leakage-safe train/validation/test split strategy
- Explain how you handle class imbalance, augmentation, and threshold tuning
- Report evaluation metrics and justify the chosen operating point
- Propose deployment and monitoring steps for production use
