Case Study
Designed and built an intelligent document processing pipeline that extracts complex policy details from unstructured commercial PDFs with 99.4% accuracy, automating 90% of manual triage.
In commercial insurance, risk assessment starts with sorting through hundreds of pages of unstructured policy declarations, loss runs, and client emails. As Head of AI Strategy & Implementation, I designed an end-to-end intelligent document processing (IDP) pipeline for Insly to automate this triage phase completely.
The architecture leverages advanced semantic chunking, dynamic prompt assembly, and secondary verification loops to prevent hallucinations. Raw documents are scanned and parsed, structured JSON is extracted, and coordinates are checked for cross-reference auditability. Hallucination checks are strictly enforced via programmatic constraints on field ranges.
This system achieved a validated production extraction accuracy of 99.4% on complex tables and text blocks. For underwriters, this reduced total submission triage time from an average of 45 minutes down to just 90 seconds, allowing brokers to issue bindings up to 10x faster while saving over €500,000 in annual processing overhead.