HGS

HGS, one of the largest insurance houses for the global shipping industry, faced an acute document processing bottleneck at the intersection of scale and complexity. Maritime insurance underwriting involves thousands of variables per policy — from vessel age and cargo type to geopolitical risk and weather exposure — making it one of the most technically demanding insurance disciplines in existence. In 2021 alone, the industry recorded over 9,000 marine insurance claims totaling $14.3 billion in damages. Extracting the structured data needed to price policies accurately from dense, multi-format insurance documents could only be performed reliably by a handful of global specialists, creating a hard ceiling on throughput, a costly dependency on scarce expertise, and growing exposure as the global merchant fleet aged past an average of 21.75 years.

HGS engaged Creative Dock to build POSEIDON, a purpose-built AI application combining object detection, computer vision, and natural language processing to automate document parsing at scale. The core NLP subsystem uses a hybrid architecture that pairs rule-based Regular Expressions with a Named Entity Recognition (NER) machine learning model built on the SpaCy Python library. Documents were labeled manually using the Doccano annotation tool, creating a domain-specific training dataset covering maritime insurance entities: plan types, insured company names, contract periods, currency amounts, and other underwriting fields. The hybrid design was a deliberate architectural choice — rules handle well-defined, consistently formatted entities, while the SpaCy NER model generalizes to the long tail of format variation that pure regex cannot cover. After training and evaluation on a held-out test set, the system was deployed into production at HGS, integrating directly into the underwriting and pricing workflow.

The POSEIDON system achieved 97% accuracy across thousands of insurance documents in production, successfully parsing the majority of the entity labels required for HGS's underwriting operations. Key outcomes include:

• 97% document parsing accuracy across all deployed document types
• Automated extraction of plan types, contract periods, insured company names, and other critical underwriting fields
• Eliminated the bottleneck created by dependence on a small pool of global specialists
• Materially accelerated insurance pricing and underwriting workflows that previously required scarce domain expertise

The application moved from development to active production use at HGS, delivering measurable operational throughput gains in a compliance-sensitive context where precision is non-negotiable.

• Hybrid architectures outperform single-method approaches: combining Regular Expressions with ML-based NER captures both structured patterns and the format variation that rules alone cannot handle.
• Domain-specific training data is the critical investment: manually labeling client documents with maritime-specific entities — not generic NLP datasets — was essential to reaching 97% precision.
• Model weight matters for production viability: lightweight NLP frameworks like SpaCy can achieve near-expert accuracy on structured extraction tasks without the infrastructure overhead of large language models.
• Scarcity of human expertise is a valid trigger for automation: when a workflow depends on a handful of irreplaceable specialists, AI extraction is both a scaling solution and a business continuity strategy.