SafeMTS
We worked with SafeMTS to deploy secure, SME-guided large language models that converted unstructured maritime near-miss reports into scalable, high-quality safety insights—cutting analysis time dramatically while unlocking predictive risk signals across 19,000+ events.
Summary
SafeMTS faced a growing challenge common to safety-critical industries: large volumes of near-miss reports rich in narrative detail but difficult to analyse consistently, quickly, and at scale. Traditional manual review processes limited the speed at which insights could be generated and constrained the program’s ability to surface systemic risks across participating operators. The objective of this engagement was to determine whether advanced AI techniques—specifically large language models—could improve analytical efficiency and data completeness without increasing reporting burden or compromising regulatory protections.
Working alongside SafeMTS and maritime subject-matter experts, we implemented secure, AI-driven workflows to transform unstructured safety narratives into structured, analysis-ready data. The solution significantly expanded the extraction of causal factors, operational activities, and high-potential risk indicators across more than 19,000 near-miss events, while reducing manual analytical effort. The outcome demonstrated a scalable model for augmenting expert judgment with AI, enabling faster insight, improved risk visibility, and a transition from retrospective safety review toward a more predictive, continuously learning safety system.
The Challenge
SafeMTS needed to extract meaningful, comparable safety insights from thousands of maritime near-miss reports submitted in free-text form by different operators, each using distinct reporting structures and terminology. Manual expert review was accurate but slow, resource-intensive, and difficult to scale, creating delays between incident reporting and actionable learning while limiting the program’s ability to surface industry-wide risk patterns.
Our Solution
We partnered with SafeMTS to deploy large language models within a secure, regulated environment, combining AI-driven text analysis with maritime subject-matter expertise. Through careful prompt engineering and iterative validation, the models were used to standardise data, populate missing fields, and extract high-value indicators—such as causal factors and high-potential events—directly from narrative reports, without increasing reporting burden on participating companies.
Results
The approach dramatically improved both the speed and depth of safety analysis, enabling structured insights to be generated across more than 19,000 near-miss events. Data completeness increased from negligible levels to near-universal coverage for key risk fields, while analytical effort was substantially reduced. The engagement demonstrated how AI can augment expert judgement, support scalable safety learning, and shift maritime risk management from retrospective review toward a more predictive, continuously improving system.