About ClaimLens
Turn accident documents into a review-ready case file.
ClaimLens is a document-intelligence demo built around the demand package for Florida auto-injury (car-accident) claims — the central artifact of a pre-litigation PI claim. It extracts the demand-package facts from messy documents, builds a timeline, flags what's missing (including the Florida PIP gates), and links every conclusion back to its source. It defaults to the law-firm framing (assemble the package, catch the gaps before you send it); the identical app also serves the adjuster / demand-package-review side.
Scope: car-accident / auto-injury claims only — the checklist, gates, and photo analysis are tuned to motor-vehicle no-fault (PIP). It's not built for slip-and-fall, workers' comp, or other injury types.
How it works
- 1
Add the documents
Open a synthetic sample case, or upload your own set — police report, medical records and bills, insurance declarations, paystub, and damage photos.
- 2
Parse into page-cited chunks
Each PDF is parsed (Docling with OCR, or a pypdf fallback) into text chunks that remember their page number — so every downstream citation resolves back to an exact page.
- 3
Embed for grounded retrieval
Chunks are embedded into Postgres/pgvector with a local model (nomic-embed-text). Retrieval keeps extraction grounded in the source text, FranchiseLens-style.
- 4
Read the damage photos
A vision model turns each photo into a structured finding — vehicle, area, severity — so multimodal evidence enters the case file, not just text.
- 5
Extract the typed schema
An LLM populates the demand-package schema (parties, injuries, costs, lost income, insurance, damages). Every field is tagged to one of the four evidentiary pillars and carries a citation.
- 6
Derive the outputs
A chronological timeline flags the Florida gates (14-day PIP rule, non-EMC $2,500 cap, 30-day insurer window); a checklist engine flags missing documents; a summary organizes everything by pillar.
- 7
Review & export
Click any conclusion to see its source page or photo, edit it (the confidence and citation stay), and download a review-ready PDF packet.
What it does that plain RAG can't
Multimodal
Reads damage photos with a vision model — not just text documents.
Structured extraction
Facts land in a typed demand-package schema, not Q&A retrieval.
Constructed timeline
Assembles dated events and flags Florida's hard date gates.
Missing-document detection
A checklist rule engine over the real demand-package list.
Auditable citations
Every fact links to its source, is editable, and shows confidence.
Privacy & data
A demo, not legal advice. The sample cases are entirely fabricated — fake names, 555 phone numbers, demo policy numbers. Don't upload real medical records or PII.
Runs fully local
FastAPI · Postgres + pgvector · Next.js. Every LLM, vision, and embedding call sits behind one adapter — free NVIDIA NIM (hosted), local Ollama (private), or a deterministic offline stub. No paid APIs.
Who built this
Alex Kargin
Data & AI engineer · Boca Raton, FL · 10+ years building production data systems
I built ClaimLens end to end — parsing, retrieval, the typed schema, the timeline and checklist engines, the review UI, and the deploy pipeline. It's one of 15 systems I've shipped, alongside FranchiseLens (document intelligence over 300+ page disclosure filings), Tennis.us (~197,000 public court records, normalized and geocoded), and LegacyToCloud (database migration with cloud DDL generation).
Open to senior data engineering, forward-deployed / solutions engineering, and applied-AI roles — remote or South Florida.