Sentinel
Automated forensic analysis of images and documents submitted with insurance claims. Detects manipulation, forgery, and fabrication — from altered receipt amounts to AI-generated damage photos.
The Scale of Document Fraud
Insurance document fraud costs the industry billions annually. Manipulated receipts, altered dates, and fabricated photos are increasingly difficult to catch by manual review — especially as AI editing tools become widely available.
Amount Inflation
Changing a receipt total from 1,849 to 18,499 to increase a claim payout. A single digit change that is invisible to the human eye.
Date Alteration
Modifying a document date to fall within a coverage or warranty period. Often done with PDF editors on scanned documents.
Document Fabrication
Adding, removing, or replacing text in scanned documents using PDF editors. White rectangles cover original content while new text is overlaid.
Photo Manipulation
Staging or digitally enhancing damage photographs for property and vehicle claims. Cloning damage, removing repairs, or compositing scenes.
AI-Generated Content
Using generative AI to create plausible but entirely fabricated documents or damage photos. Modern AI tools make this increasingly accessible.
Copy-Paste Forgery
Copying and pasting regions within a photo to duplicate or conceal damage. Used to exaggerate the extent of property or vehicle damage.
Three-Layer Forensic Architecture
Sentinel uses a layered approach aligned with industry best practice for document fraud detection. Every submitted file passes through all three layers automatically, and the results are cross-correlated to produce a single risk assessment with specific, actionable findings.
Signal Analysis
Over 40 automated analyzers examine the file at pixel level. Each analyzer is based on published forensic research and looks for a specific type of manipulation — from compression artifacts and noise inconsistencies to copy-paste traces and camera sensor fingerprints. These run in parallel, typically completing in 1-2 seconds.
Metadata Analysis
The file's embedded metadata is extracted and cross-validated. This includes camera information, timestamps, editing software traces, GPS coordinates, and document structure. For PDFs, Sentinel detects incremental edits, embedded fonts in scanned documents, and content overlaid after scanning.
AI Semantic Analysis
Two AI vision models review the document in sequence. The first reads every number and verifies the arithmetic — does unit price times quantity equal the line total? Does the sum add up? The second cross-correlates all evidence from every layer and delivers the final judgment.
Research-Backed Signal Analysis
Every automated analyzer in Sentinel is grounded in peer-reviewed forensic research. The system doesn't rely on a single detection method — it runs dozens of independent checks simultaneously, each looking for a different trace of manipulation.
- Compression artifacts reveal regions saved at different quality levels
- Sensor noise patterns identify content from different cameras
- Copy-paste detection finds cloned regions within an image
- Frequency analysis detects AI-generated or GAN-produced images
- Font and text alignment checks catch pasted text in documents
- Camera sensor fingerprinting verifies all regions were captured by the same device
PDF Document Forensics
PDF documents require specialized analysis beyond pixel-level forensics, as manipulation often occurs at the structural level. Sentinel renders each page and compares the result against the embedded scan — any difference reveals content added after scanning.
- Scan-vs-render comparison detects overlaid text and white-out rectangles
- Font inventory flags embedded fonts in documents that should be pure scans
- Structure analysis detects incremental saves indicating post-scan editing
- Embedded images are individually extracted and analyzed for manipulation
- Multi-page cross-checking compares dates, names, and amounts across all pages
AI Vision Verification
Signal analyzers provide objective, reproducible measurements. But some fraud requires understanding: does the arithmetic add up? Does the date make sense? Is the document internally consistent? Sentinel's two-stage AI vision pipeline addresses this.
- Reads every number on receipts and invoices, then verifies all arithmetic
- Checks tax rates, currency formats, and merchant details for consistency
- Compares dates, reference numbers, and amounts across multiple pages
- Applies Benford's Law to check whether financial amounts follow natural distributions
- Cross-correlates all signal findings with visual evidence for a final judgment
- Identifies false positives and explains its reasoning
Deep Learning Forgery Detection
Beyond traditional signal analysis, Sentinel uses purpose-built neural networks to detect manipulations that leave no visible trace to the human eye.
- Mesorch (AAAI 2025) combines CNN and Transformer architectures to produce pixel-level forgery maps for spliced, cloned, and inpainted regions
- DINOv3, a custom forensic model built on a 6.7 billion parameter vision backbone, is trained specifically to detect AI-generated inpainting — the type of manipulation that defeats most commercial detection tools
- Trained on specialized datasets that prevent the model from learning processing shortcuts, forcing it to identify actual manipulated content rather than image processing artifacts
- Produces visual heatmaps highlighting exactly which regions are flagged, with confidence scores for each
Proven Detection Accuracy
Sentinel has been tested against real-world fraud scenarios. Genuine documents receive low risk scores; manipulated documents are flagged with specific evidence.
| Test Case | Risk Score | Key Finding |
|---|---|---|
| Clean receipt photograph | 6.8 / 100 | Arithmetic checks out correctly |
| Forged receipt (amount changed) | 100 / 100 | Line total inconsistent with unit price and quantity |
| Clean PDF scan (3 pages) | 6.8 / 100 | No manipulation detected |
| Forged PDF (date changed in editor) | 100 / 100 | Date mismatch across pages, 8 fonts found in scan |
Two Ways to Integrate
Sentinel fits into your existing claims workflow. Choose the integration method that works best for your organization.
REST API
Integrate Sentinel directly into your claims processing system. Submit documents via API and receive structured forensic reports — pushed to your endpoint via HMAC-signed webhook on completion, or fetched on demand.
- Submit jobs with optional
webhookUrlandwebhookSecret— receive results via HMAC-SHA256 signed webhook callbacks - Manage delivery with retry and cancel endpoints (
POST /sentinel/jobs/{guid}/webhook/retry,DELETE /sentinel/jobs/{guid}/webhook) - Polling fallback via
GET /sentinel/jobs/{guid}for synchronous workflows - WebSocket support for live progress updates
- Reports available in 16+ languages
- Visual heatmaps and annotated images returned as URLs
- Supports JPEG, PNG, HEIC, TIFF, WebP, GIF, BMP, and PDF
Integrated in LENS
Use Sentinel through LENS, our web-based investigation platform. Claims handlers can upload documents directly in the browser and review forensic results with interactive heatmaps and highlighted findings — no technical integration required.
- Upload and analyze directly in the browser
- Interactive forensic heatmaps and annotations
- Plain-language forensic explanations for claims handlers
- Combined with credit reports, property data, and company information
- No developer resources needed — ready to use immediately
Comprehensive Format Support
Sentinel analyzes all common image and document formats submitted with insurance claims. JPEG files receive the deepest analysis including compression-specific forensics. PDF documents are rendered, compared against embedded scans, and reviewed across all pages for cross-page inconsistencies. HEIC files from iPhones are fully supported.
- JPEG — Full pixel forensics including compression analysis, quality table fingerprinting, and ghost detection
- PNG, TIFF, WebP, GIF, BMP — Pixel forensics, noise analysis, and AI detection
- HEIC — Native support for iPhone photos with full forensic analysis
- PDF — Page rendering, overlay detection, structure analysis, font inventory, multi-page AI review
Ready to Detect Document Fraud?
Reduce fraudulent claim payouts with automated forensic analysis. Available as API or integrated in LENS.