Banuba published a walkthrough of face detection in KYC onboarding that makes an argument worth repeating: the expensive part of identity onboarding is not the verification, it is the round trip. An applicant photographs themselves, uploads, waits, and gets rejected for glare or a bad head angle. Banuba’s position is that the capture quality check belongs on the phone, before submission — the SDK reads more than ten facial parameters locally, including head position, distance from camera, skin tone and expression, so the failure is caught while the user still has the camera open. They quote 60 FPS on mid-range mobile hardware across a -90° to +90° head-angle tracking range, against a baseline of iOS 13+, Android 8.0+, and a 1280x720 camera at 30 FPS minimum.
The second claim in the post is the one with regulatory teeth. Face tracking and analysis run on the device and no user data is stored, which means the biometric template never becomes a cross-border transfer problem. That is a materially different compliance posture from an API that ships a face to a server for scoring, and it is the reason a capture-quality gate and a liveness check are worth separating in your architecture rather than collapsing into one vendor call.
The catalog carries Banuba with three published surfaces, and the split matches the article’s argument closely. The Face AR SDK is the tracking and landmark layer that does the capture-quality work described in the post. The Face Liveness SDK is the separate anti-spoofing surface that verifies a face is real and present — the piece that actually carries the fraud burden in a KYC flow. The Video Editor SDK is the third and belongs to a different job entirely. Banuba’s Kin Score sits at 23.7, emerging, on strong discoverability (72.2) held down by contract quality of 10.7 — the predictable shape for an SDK company whose surface is native libraries rather than HTTP contracts.
The detail we would flag for anyone reading the post as an integration plan: Banuba’s Agent Readiness score is 6.1, and exactly one dimension is lit. They publish agent skills — thirteen of them are indexed in the catalog — and nothing else. No spec presence, no documented auth clarity, no error semantics. For a human developer following the quickstart that is fine. For a coding agent asked to scaffold a KYC capture flow, publishing skills without publishing a contract means the agent gets instructions it cannot verify against anything. The skills are the right instinct arriving ahead of the foundation.