Case study · 2023 – 2026
Automated Vehicle Damage Assessment
Ensemble computer-vision system that detects vehicle damage and missing parts for insurance claims - 90% damage-detection accuracy.
90%
Damage-detection accuracy
45%
Lower manual inspection costs
30%
Lower annotation costs per year
Context
XA Group (Xpress Automation), motor-insurance claims, 2023 – 2026. I led full-stack development of the deep learning systems behind automated claim assessment.
Problem
Assessing a motor-insurance claim needed a manual inspection of the vehicle, which made claims slow and costly. Automating it required models that cover many damage types - and labelling enough data for every new damage type was itself expensive.
Approach
An ensemble of computer-vision models detects and localises damage and identifies missing parts from claim photos, so claims can be assessed without a manual inspection.
Self-supervised pre-training on unlabelled claim images cut the number of annotated examples needed for each new damage type. Models were optimised with TensorRT for inference.
After launch I owned the models in production, monitoring for data drift and retraining when the incoming distribution shifted (AzureML, MLflow).
Results
- 90% damage-detection accuracy
- 45% reduction in manual inspection costs
- 30% annual reduction in annotation costs via self-supervised pre-training
- Drift monitoring and retraining with AzureML and MLflow
Stack
- PyTorch
- Hugging Face
- Ultralytics
- TensorRT
- AzureML
- MLflow
- Docker
Have a similar problem?
Tell me what you're working on and I'll tell you honestly whether I can help.