Transforming Business with AI

Optimized defect detection model from 75% to 91.2% precision, reducing false positives by 63%.

Challenge

The client had trained a defect detection model using visual inspection footage, but performance had plateaued at ~75% precision. False positives were creating costly interruptions in production lines.

Solution

• 1
  Conducted a full audit of data preprocessing, model architecture (ResNet-50), and evaluation pipeline
• 2
  Identified underrepresented defect classes and mislabeled samples
• 3
  Introduced weighted loss functions and focal loss to improve sensitivity

Execution

• Ran k-fold cross-validation to validate generalization
• Improved label quality through stratified relabeling efforts
• Tested 3 model variants with hyperparameter tuning and augmented image datasets

Results

Key Metrics