Claim Review
An AI helper for insurance claim reviewers: it reads a claim's documents, checks them against the company's own rules and drafts a preliminary evaluation that a person then decides on.
OngoingStarted November 2026
What it is
Claim Review is a multi-tenant product for insurance companies. A company uploads its rules and procedures, optionally per insurance product. A reviewer creates a claim and uploads its documents, PDFs and photos, then presses Evaluate. The AI does the reading and the first check. A person always makes the decision and records it.
What a reviewer gets
- Extracted fields with page references, so every value can be checked against the source page.
- Findings that cite rule clauses, instead of a verdict with no reasons behind it.
- Missing documents, a reimbursement estimate computed by code rather than by the model, review flags, a recommendation and a draft request.
- A chat about one claim, grounded only in the rules and that claim's documents.
Personal data first
Documents are read locally, from the text layer or with Tesseract OCR, and personal data is redacted before any cloud model sees the text. Team members never see personal data or the original claim files.
The console
Super admins create teams with their first admin and run the LLM side of the product: tokens, cost, traces, platform guardrails and evals.
How it is built
A Turborepo monorepo with a NestJS GraphQL API, two TanStack Start apps and a Python AI service. FastAPI and an arq worker run LangChain and LangGraph with Claude Haiku, with Voyage embeddings and rerank over pgvector. It has not launched yet; the screenshots show development data.
