Verafica detects AI-generated images with calibrated confidence, layered detection signals, and evidence you can audit — built for how images actually arrive.
No single signal is a single point of failure. Verafica layers independent detection streams and routes disagreement to human review.
A CLIP ViT-L/14 backbone with a calibrated classification head reads the statistical fingerprint of the image. Scores are calibrated probabilities with decision thresholds — not vibes.
A vision-language witness inspects image content for structural and semantic impossibilities — garbled text, impossible geometry, anatomical errors. It annotates; it never decides alone.
C2PA Content Credentials and EXIF camera data are verified first. A cryptographically valid AI signature settles the question before any model runs.
Pixel-level heatmap of manipulated regions for edited real photos — splices, object removal, AI inpainting.
Face-swap and identity-manipulation detection for photos and stills.
A single REST endpoint. Embed verification in claims queues, upload flows, and onboarding pipelines.
Six decision tiers. Uncertain images are flagged for review — never silently auto-cleared or auto-condemned.
Every verdict ships with its evidence clauses, signal breakdown, and model version. No black boxes.
Verdicts are re-checked under JPEG re-compression, resizing, and cropping — the edits real uploads go through.
Pilot uploads are analyzed in memory and not retained. Storage only ever happens with explicit consent.
Measured robust to WhatsApp and social-media compression — the way images actually arrive in India.
AI fake profiles, NRI fraud, pig-butchering via apps. Synthetic profile photos are the entry point.
High-volume consumer fraudFraudulent claim photos, manipulated damage evidence, synthetic documents attached to claims.
Claims-desk exposureFabricated exhibits and AI-generated imagery entering evidence chains that assume photos are real.
Chain-of-custody riskSynthetic KYC photos, altered income documents, forged statements on loan applications.
Onboarding-fraud exposureSee what our detection engine says about the images crossing your desk — with calibrated confidence and evidence attached.