NezT Labs: AI Photo Fraud Triage

NezT Labs (in-house R&D) · AI / Computer Vision

An in-house build: a batch ML pipeline that flags visually inconsistent images in field-verification photo archives, with no labelled training data required.

How it works

DINOv2 embeddings (with an optional CLIP adapter) compare each outlet only against its own photo history.

Three fused signals (robust centroid distance, kNN consensus, Isolation Forest) produce one suspicion score.

Per-outlet adaptive thresholds, content-addressed caching and reproducible, audit-ready outputs.

Tech stack

Python, PyTorch, DINOv2, CLIP, scikit-learn, Pydantic, Typer, Docker

Impact

Triage of large field-photo archives without humans reviewing most images. No labelled data and no model training needed.