DAFNE
DAFNE (Digital Anti-Fraud Neural Engine) is VISADA’s AI engine that analyzes accident images to identify patterns of damage across different cases. The system does not assign abstract scores or probabilistic indicators: it builds a technical representation of the damage and compares it with the company’s history, producing verifiable visual evidence. The process is designed to be continuous, scalable, and traceable.
Preparation
Construction of the historical base (one-off)
To identify previous claims, it’s necessary to start with historical data. At the start of the collaboration, DAFNE analyzes a set of previous claims deemed relevant by the company, typically two to five years of images. The system applies the same pipeline steps to each image: it anonymizes the content, identifies and isolates the damage, analyzes its shape, extent, and type, and builds a representation of the damage comparable over time. In this way, each claim becomes an indexed unit within the VISADA database, creating the initial baseline for recurrence detection. Despite the high volumes, processing is completed in just a few days thanks to a batch workflow optimized to work in parallel on large quantities of images. At the end of this phase, the system is ready to detect fraud.
Operational Analysis
Daily use
Ingestion
of claims
On a daily basis, the company sends mass claims to be verified according to its own rules (value, geographic area, sampling, alerts).
Automatic
anonymization
The images are stripped of sensitive elements (license plates, faces), making the visual data anonymous: what matters is the damage, not the identity.
Damage
Analysis
DAFNE processes the images by applying a structured pipeline: it identifies and isolates the damage, analyzes its shape, extent and typology and builds a representation of the damage comparable over time.
Matching
with the historical
The detected damages are compared with those in the historical database to identify recurrences or reuses between different claims.
Human
validation
High-confidence matches are verified by specialized analysts to eliminate false positives and confirm relevant cases. This step ensures operational reliability without compromising the system's scalability: human intervention is focused only where necessary.
Report
Generation
If confirmed, a structured evidence report is produced, which includes the pair of damages being compared.
Output
Operational Evidence Report
The output produced by DAFNE is not a simple alert or an abstract risk score, but a report of evidence designed for operational use. The report provides a direct visual comparison between the incident being analyzed and related cases identified in the history. Areas of damage are highlighted and correlated through targeted image clippings, making the match immediately understandable. Each finding is accompanied by clear identification and timestamps, ensuring traceability and verifiability. This way, anti-fraud and claims settlement teams receive documented and actionable output, reducing the need for further preliminary checks. Thanks to its visual and historical nature, the report can also be used in more formal contexts, with the aim of accelerating decisions and interventions.
