Automating damage assessment and boosting operational efficiency for a car insurance company

Any successful project starts with a structured and well-defined approach. To address the client's challenges, we designed a comprehensive strategy that ensured a seamless transition from manual car damage assessments to an AI-driven, automated process.

• Step 1: Assembling the right team

We carefully selected a team of experts tailored to the project's needs, including an AI/ML Engineer, two React Native Developers, two Backend Developers, a QA Engineer, and a Project Manager.

• Step 2: Structured planning & R&D

We initiated the project with extensive research and development (R&D) to analyze existing insurance data and damage assessment workflows, develop, train, and validate AI/ML models capable of identifying six distinct damage types: dent, scratch, crack, glass shatter, lamp broken, and tire flat. Also, we needed to define a roadmap for integrating the AI-powered assessment process into a mobile app.

• Step 3: Agile execution

We broke the project into iterative sprints using an Agile SDLC methodology, allowing for continuous feedback and refinements. Key development milestones included:

- Building a mobile application that minimizes manual intervention and speeds up claim approvals

- Developing backend services to support secure and scalable data processing

- Implementing AI-powered automation to classify car damages and estimate repair costs with high accuracy

• Step 4:

Our collaboration with the client is ongoing, ensuring continuous enhancements, model optimizations, and seamless integration into their broader insurance ecosystem.

Multi-Stage AI Model Training

We implemented a structured, multi-stage AI/ML training approach to enable precise car damage detection and classification:

• Stage 1: Car part segmentation training

- Utilized a segmentation network with a ResNet backbone and FPN to detect and delineate car parts.

- Trained on a specialized car parts dataset to recognize different vehicle components with high precision.

• Stage 2: Damage detection training

- Applied transfer learning from Stage 1 to a new model trained on a car damage dataset.

- Integrated a Region Proposal Network (RPN) to detect damage locations.

- Introduced a damage type mask network to classify six types of damages (dent, scratch, crack, glass shatter, lamp broken, tire flat)

• Stage 3: Inference & real-time damage assessment

When a new car image is uploaded, the model first identifies car parts, then detects damages within those parts.

The model associates damage areas with specific car parts using segmentation mask overlaps.

Data aggregation assigns additional metadata (car model, year, severity level) for improved insurance assessment.

Seamless Integration with a Scalable Tech Stack

To ensure smooth performance and scalability, we built the solution using the following technologies:

• Mobile app development:

React Native – Cross-platform mobile app for policyholders and brokers

Node.js – Backend logic for processing claims and approvals

• AI/ML development:

Python – Core language for AI/ML model development

AWS Sagemaker – Model training, fine-tuning, and deployment

• Cloud & Infrastructure:

AWS CloudFront, AWS EKS, AWS Route53 – Ensuring reliability, scalability, and security

AWS CodeCommit, CodeBuild, CodeDeploy – Enabling CI/CD automation for efficient updates

Automated Approval & Process Optimization

The mobile app enables instant damage detection and cost estimation, allowing:

• Minor damages to be automatically approved for repair, reducing broker workload
• Insurance employees to review aggregated and labeled damage data with precise severity levels
• Customers to receive faster claim approvals, enhancing satisfaction

By leveraging advanced AI/ML capabilities, Binariks delivered a highly efficient, automated car damage assessment solution that optimizes operational processes while improving user experience.