Claimsight

Underwriters were spending their expertise on data entry, not risk. An MGA handling a large book of commercial property business had underwriters manually opening every incoming submission (PDFs, spreadsheets, loss runs, SOVs) and re-keying the same fields into their policy system before they could even start evaluating the risk. A single complex submission could take hours just to get decision-ready, and audits of past files meant repeating that process file by file, program by program.

The team didn't need another dashboard. They needed the busywork gone: extraction, consistency checks and comparison done for them, with underwriters spending their time on judgment calls instead of document wrangling.

A pipeline from raw document to decision-ready file. Claimsight ingests submission documents in whatever form they arrive (scanned PDFs, Excel SOVs, emailed loss runs) and turns them into structured, queryable data. I built the system end to end: the document-processing pipeline, the extraction and validation layer, the underwriter-facing review app, and the audit and comparison tooling layered on top.

Documents are OCR'd and parsed into structured fields (location, construction type, occupancy, prior losses, coverage limits), cross-checked against the firm's own underwriting guidelines, and flagged wherever something is missing, inconsistent, or outside appetite. Underwriters get a single decision-ready view instead of a stack of PDFs, and every flagged item shows exactly which source document and field it came from.

Extract with confidence scores, never with silent guesses. The core architecture separates extraction from judgment. A document-processing service turns unstructured files into structured records with a per-field confidence score. A guideline engine checks those records against the firm's own underwriting rules, not a generic model's assumptions, and only high-confidence, rule-passing fields are auto-populated; everything else is routed to a human review queue with the source document highlighted next to the field in question.

This kept the system auditable by design: every extracted value traces back to a page and a bounding box, every flagged inconsistency traces back to the specific guideline it violated, and nothing was decided by the model, only surfaced for a human to decide on.

Designed for a regulated industry that audits everything. Because underwriting decisions carry compliance weight, I treated traceability as a first-class feature, not an afterthought. Every extraction, flag and comparison is logged with its source and confidence score, so the firm's own audit process could verify the tool's output the same way it audits an underwriter's file. Rate limiting and retry logic on the OCR and LLM pipeline meant a malformed PDF or a provider hiccup degraded to a manual-review flag instead of blocking the queue.

Adoption followed a familiar pattern: underwriters were skeptical of anything that touched their file review at first, but once they saw the source-linked flags matched what they would have caught by hand, just faster, the tool became the default first pass on every new submission.

Submissions arrive decision-ready instead of as a stack of PDFs, and the audit process can verify the tool's output the same way it audits an underwriter's own file.