A valid QR proves the certificate is real. Visual authentication proves the object in the buyer's hands is the one that was certified. Consumers photograph the product on /verify and Fidaty matches it against the producer's reference images with CLIP embeddings — a second, independent layer of trust.
2 layers
Of trust
Certificate + object
512-dim
CLIP embedding
Per reference view
0
Photos stored
Embedded then discarded
Tier-driven
Match thresholds
Tunable, no redeploy
Counterfeiters clone QR codes and links. Visual authentication closes that gap: it ties the digital certificate to the visual identity of the physical product, so a cloned code on a fake item fails the photo match.
The producer uploads one or more reference views (front, back, serial detail) for a certificate. Each is embedded with a CLIP-class vision model and stored as a 512-dim vector — never as a raw image a verifier can lift.
Front, back and serial-detail views make the fingerprint robust and far harder to spoof with a single photo pulled from the web.
A vision model turns each reference into a 512-dimensional vector on pgvector. Pluggable provider (HuggingFace / custom) with clean degradation.
Reference images live in a dedicated bucket, write-restricted to the certificate's owner via row-level security.
On /verify, the consumer takes a live photo. Fidaty embeds it server-side and compares it to the certificate's reference views with cosine similarity — returning a confidence band, never a raw yes/no.
On mobile the camera opens directly; the image is downscaled before upload for speed and privacy.
The live embedding is matched against the certificate's reference vectors. The best score drives the verdict.
The submitted image is embedded and discarded. Only the verdict, score and coarse geolocation are kept.
Visual authentication returns strong / weak / no match — deliberately a confidence band, because a binary 'fake' carries legal weight the platform won't own. Failed matches also feed the counterfeit radar.
Tier-driven thresholds map cosine similarity to a clear, honest confidence band shown next to the on-chain proof.
Every geolocated no-match becomes a signal for the cross-signal counterfeit cluster radar.
Without a configured vision provider the feature simply hides — no broken UI, no false confidence.
A buyer on the secondary market photographs a watch or handbag and confirms it matches the brand's reference images before paying.
An installer checks that a safety-critical component visually matches the certified part, not a look-alike.
A gallery confirms the physical piece matches the artist's enrolled references, beyond the certificate alone.
Packaging is matched visually to catch refilled or repackaged counterfeits that clone the QR.
Every service in the Fidaty platform is integrated with the rest: identity, compliance, analytics and consumer engagement flow seamlessly. No data silos, no duplicate work.
Book a demoEnroll reference photos per certificate and let consumers confirm the physical object on /verify.