Skip to content
All projects
Case studyCS/12
StatusPersonal project
TypeAI
Year2026

ClaimSense

An LLM reads the insurance claim; plain code decides where it goes

Upload a First Notice of Loss as a PDF or text file. pdfplumber extracts the text, an LLM on NVIDIA NIM fills a Pydantic schema, and a deterministic rules engine routes the claim to fast track, manual review, the specialist queue or investigation, with a written reason.

GitHub
  • 4 routing outcomes
  • Temperature 0
  • Schema-validated output
  • 5 req/min per IP

§1

The problem

Claims handlers read every FNOL document to pull out the same fields and choose a queue. LLMs are good at reading messy documents, but a routing decision has to be explainable and repeatable.

§2

My role

Sole developer: document parser, extraction agent, routing engine, rate limiter and the single-page dashboard.

§3

Architecture

Fig. 1 · ClaimSense architecture
textvalidatedFNOL UPLOADPDF · TXTRATE LIMITERsliding windowPARSERpdfplumberEXTRACTORPydantic schemaNVIDIA NIMLLM · temp 0RULES ENGINEdeterministicROUTE + REASON4 outcomes
clientserviceexternal

§4

Key decisions

  1. D/01

    AI extracts, code decides

    The LLM only fills in the schema. Routing is ordinary Python that a reviewer can read, test and audit.

  2. D/02

    The schema is the prompt

    The Pydantic model's JSON Schema is sent with the document at temperature 0, and the reply is validated with model_validate before anything uses it.

  3. D/03

    Rules in priority order

    Missing mandatory fields → manual review; fraud words (staged, suspicious, inconsistent…) → investigation; any injury → specialist queue; under $25,000 → fast track; otherwise manual review.

  4. D/04

    Protect the LLM quota

    A sliding-window middleware allows 5 requests a minute per IP on /api/process and 30 elsewhere, returning 429 beyond that.

§5

Challenges

  • Models wrap JSON in markdown fences despite instructions, so the agent strips them before parsing and fails clearly when the output still isn't valid.
  • Injury detection treats "none", "no" and "n/a" as no injury, so the way an LLM writes "nothing" doesn't send a claim to the specialist queue.
  • Processed claims are kept in localStorage, so the dashboard keeps its history without a database.

§6

Stack

  • Python
  • FastAPI
  • Pydantic
  • NVIDIA NIM
  • pdfplumber
  • JavaScript

Next case study

Procuro

RFPs written from one sentence, vendor replies parsed from email, bids compared