Claims Document Agent

A synthetic auto bodily-injury claim, ready to process

11-page packet: FNOL, police report with a crash diagram, ER and imaging records, a physical-therapy course, itemized bills, and a demand letter. Click Process this claim to run the extraction & validation pipeline.

📄 Source packetPage 1 / 11

🖼 Extracted image Layout Parser

Diagram detected in the police report and attached to the narrative, tied to its source page.

Crash diagram · rear-end impact, both vehicles northbound p.2

Key fields Custom Document Extractor

Predefined fields pulled from the packet. Each carries a source-page citation and a confidence score — click a citation to jump to the page.

Medical chronology date-ordered

Events ordered by the date parsed from the record text (not ingest time), each with provider, source page, and confidence.

Business values & reconciliation

Monetary values extracted, then reconciled — the itemized charges are summed and checked against the stated total.

🛡 Validation agent human-in-the-loop

A separate agent scores the output and triages what a human must review before settlement — it does not decide.

Generated narrative high-fidelity grounded

Composed against business rules, grounded strictly in the packet. Every claim carries an inline citation.

On GCP this pipeline is Cloud Storage → Document AI (OCR + Custom Extractor + Layout Parser) → Vertex embeddings + Vector Search → Gemini (high-fidelity grounding) → Gen AI Evaluation Service, orchestrated with a durable workflow, wrapped in VPC-SC + CMEK + audit logging. This prototype runs the same logic on a synthetic packet.