AI-Generated Image Detection for Insurance Claims
Vladimir M., insurance fraud detection proof of concept
Delivered a proof of concept that scores uploaded claim photos for the likelihood they were AI-generated, using the CO-SPY detection method (CVPR 2025), deployed on AWS with a shareable demo.
The Challenge
Generative image tools make it easy to fabricate or alter photos submitted as claim evidence. Vladimir needed a fast proof of concept to see whether AI-generated images could be flagged automatically, starting from photos taken on ordinary phones. The original brief covered video; early testing showed still images were the stronger, more immediate use case, so the scope shifted to images.
Our Solution
We built an end-to-end proof of concept around CO-SPY, a detection method that combines semantic features with pixel-level artifacts so it holds up across multiple image generators. Users upload a JPEG or PNG through a simple Next.js web app; a FastAPI backend queues the job with Celery and Redis, runs the PyTorch model, and stores uploads in S3. The stack is containerized with Docker Compose and deployed on AWS EC2, so stakeholders could test it from a shareable link.
Results
Vladimir got a working, shareable demo that returns a 0–100% AI-likelihood score with clear threshold bands (not AI-generated, likely AI-generated, very likely AI-generated), plus a deployable architecture to build on. In his words: “I hired them for a small project, and I couldn't be happier!”
Have a similar problem?
Talk it through on a free 30-minute call, or send a short note first.