AI-Augmented Severity Evaluator for Port State Control (PSC)
The evaluation of Port State Control (PSC) inspections in the maritime industry is highly tedious and subjective to the inspector, resulting in operational inefficiencies or critical oversights that cause accidents. To address this, we developed an AI-augmented severity evaluation tool that provides an objective "second opinion" for PSC findings, enhancing both reliability and scalability of vessel inspections which landed us 1st Place in the Maritime Hackathon 2025.
📊 Synthetic Data Generation and Preparation
The solution integrates GenAI (Google Flan-T5) amd automated ML (AutoGluon) to generate synthetic data for each inspection. The process begins with structured data preparation, extracting key features from inspection reports. Google Flan-T5 was prompt engineered with CoT reasoning to generate unbiased severity ratings for each finding - mirroring the reasoning of expert inspectors but free from individual bias. Each inspection will result in 1-4 inspectors' severity rating along with 1 synthetic rating to create a more balanced and representative dataset.
🤖 Machine Learning Model
Built with Autogluon, the machine learning model is trained with the augmented dataset, capable of accurately evaluating the severity of new inspection findings. The best-performing model was CatBoost with a ROC AUC score of 0.768, demonstrating strong predictive performance.
🌟 Key Features 🌟
- AI "Second Opinion" for Data Augmentation to improve model generalizability.
- Consensus scoring through aggregation of human and AI scores for a balanced, industry-aligned severity evaluator.
- Seamless integration with PSC workflows, supporting fair and efficient vessell inspections at scale.