Projects

DAFNE — Digital Anti-Fraud Neural Engine

DAFNE is Visada’s flagship AI platform to detect repeated or manipulated vehicle damages across insurance claims. It leverages state-of-the-art computer vision models and a microservice cloud infrastructure to reduce fraud and ensure fairer insurance.

Problem

Insurance companies face billions in losses every year due to fraudulent claims.
Typical fraud involves reusing the same vehicle damage photos across different claims, or manipulating images to exaggerate costs.
Manual detection is slow, expensive, and highly error-prone.

Solution

DAFNE (Digital Anti-Fraud Neural Engine) is an AI-powered platform designed to detect repeated or tampered damages across claims by analyzing vehicle images at pixel and geometry level.
It ensures more transparent, automated, and fair claim handling, with a human-in-the-loop validation for sensitive cases.

Key outcomes:

  • Fraud reduction across the claims lifecycle.
  • Cost savings for insurers.
  • More equitable premiums for honest customers.

Architecture

DAFNE is built as a cloud-native microservices platform:

  • Ingestion Service → processes uploaded claim photos and normalizes metadata.
  • Damage Extractor → detects damages and generates embeddings via custom ViT backbones.
  • Vector Search Service → Qdrant DB for fast similarity search on millions of embeddings.
  • Matching Engine → checks geometric consistency.
  • Report & Audit Service → produces explainable results and integrates with insurer workflows.

All services are containerized with Docker, orchestrated on Azure Kubernetes Service (AKS), and communicate via REST APIs and a message broker.

Security: Teleport for access, RBAC policies, and ISO 27001-aligned practices.

Tech Stack

  • AI → PyTorch Lightning, ViT, DINOv2, Triplet Loss, W&B tracking.
  • Backend → Python/FastAPI microservices, PostgreSQL, Qdrant vector DB.
  • Infra → Kubernetes (AKS), Docker, GitHub Actions CI/CD, Signoz monitoring.
  • MLOps → Weights & Biases for lineage, Paperspace for training, TorchScript export for inference.

Results

  • Detects duplicate or manipulated damages in real time.
  • Processes millions of images with scalable vector search.
  • Generates auditable, explainable evidence for insurers.
  • Currently piloted with major Italian insurance companies.

Status

DAFNE is actively developed by Visada Srl, an Italian insurtech startup, and is being deployed in pilot projects with top insurers.