VeriDoc — Insurance Document Fraud Detection

Overview

VeriDoc helps insurance investigators triage claim documents such as medical bills, invoices, prescriptions, identity papers, vehicle documents, repair estimates and police reports. It never declares a document genuine or fraudulent. Every finding carries a severity, a calibrated confidence, the measured evidence and the detector version, and the final claim decision is always recorded against a named human.

How it works

Uploads go through a 14-stage analysis pipeline on background workers:

• PDF structure: metadata, revision history, producer software, data appended after the end-of-file marker, and embedded JavaScript.
• Image forensics: error-level analysis, sensor-noise residuals, copy-move detection, JPEG-ghost compression history and resampling checks. Region significance comes from a sliding-window statistical test.
• Typography: baseline, size and stroke analysis. For digital PDFs it reads the exact font of each character.
• OCR and extraction: text from the PDF layer or Tesseract, with a bounding box for every field.
• Rules: 20 deterministic checks, including line-item arithmetic, tax rates, date logic, VIN check digits, IBAN mod-97 and duplicate content across claims.
• Cross-document comparison: fields are compared within a claim bundle using type-aware pairing policies.
• Synthetic media and trusted sources: both sit behind provider interfaces.

Architecture

A Next.js 15 investigator console talks to a FastAPI backend through an httpOnly-cookie proxy. RQ workers process the queue, PostgreSQL stores 25 tables, and S3-compatible storage keeps originals immutable and separate from derived artifacts. Four roles (investigator, claims officer, reviewer, administrator) are enforced by the API, and every action is audited. A deterministic seeder builds a demo with 10 claims and 22 generated documents that contain real, detectable anomalies. The backend has 174 automated tests.