Oz Forensics

Presentation attack and deepfake detection, one to one and one to many face matching, and document recognition are what it sells, delivered as kits, a hosted service and on premises licenses. No workflow, case manager or screening database sits beside the models, so removing them leaves nothing.

Each liveness check returns a determination with no settable confidence threshold, escalation route, manual review surface or attempt limit published. An institution consuming the signal inherits an outcome it did not produce.

Full BixeLab reports for presentation and injection attack detection and biometric performance sit in the trust center, alongside ISO/IEC 19794 5 image quality assessments and iBeta results. That is test documentation, not model documentation, and no validation summary or drift monitoring is described.

Alvos's chief technology officer says verifying a customer went from hours to minutes, and Caina and Unico describe guidance and reliability gains, beside more than 3 million authentications a day. The Alvos change carries no figure or method, and laboratory results measure the technology rather than customer outcomes.

Model improvement is limited to customer data explicitly provided for it, and no pooled dataset or shared negative list is described. What counts as explicitly provided data is left for contract.

The trust center names a data privacy officer and covers data subject requests and breach notification, with GDPR compliance stated. Retention for facial images and liveness video on the hosted path is unpublished, and Illinois, Texas and Washington biometric law is not addressed.

ISO/IEC 27001:2022 and a SOC 2 Type 2 report, both from 2025, and a Level 1 Cloud Security Alliance STAR entry are presented as artifacts in the trust center. Certificate numbers and the SOC 2 audit period are not shown.

Technical conformity assessments against biometric standards are its credentials, not regulatory authorizations, and no financial license is stated. Government bodies appear as a customer type without any agency or program named.

Its own face matching model was submitted to NIST's face recognition evaluation, which measures demographic differentials by design, and a 99.99 percent accuracy figure from it appears on the homepage. Only that headline number surfaces, not the demographic breakdown, and performance in emerging market capture conditions is not analyzed.

Liveness rejections usually look to the person like a failed attempt with a prompt to retry, so someone repeatedly refused may never learn a determination was made. Beyond a data subject request route, Oz publishes no accuracy guarantee or review right, and responsibility across Oz, the identity platform and the institution is unallocated.

Subprocessors are named in the trust center: Google Cloud Platform in the United States for model development, Amazon Web Services and three business tools. Its models are its own, though no model or version identifier is given to record against a decision.

Web and mobile SDKs with front end customization and on device or server execution make integration the delivery mechanism, and Unico embeds the technology in its own product. Which bank onboarding stacks it has been proven in is not stated.

Hosted, on premises and a no storage Instant API are offered, with liveness able to run on the user's device, so biometric data need never leave the institution. For the hosted service, region and transfer mechanism outside the United States are not given.

A hosted service, on premises licenses and AWS and Huawei marketplace listings show how a deal can be shaped, but no rates, tiers or trial terms are published. Neither the per check rate for hosting nor the license basis for on premises use is stated.

More than 30 countries, weighted toward emerging markets, are reached across banks, fintechs and identity platforms, among others. As a component supplier, much of its institutional reach runs through other identity platforms, and material is organized by product.