Hi HN—Berwin here.

We built ExtractQ after watching friends in insurance re-type PDFs all day. The service uses a vision-plus-LLM pipeline (AWS Bedrock + LangChain + CrewAI) to map any incoming doc to JSON without per-form training. We then validate via third-party APIs (DMV, VIN) before ingesting into the core claims app.

In production at a mid-size insurer we cut average claim prep time from 45min to 19min and reduced downstream data-entry errors by 85%.

I’d love feedback on scaling doc-heavy AI workflows—especially around validation and audit trails.