Pipeline. Every frame: our face detector (FCOS-style, 100K params, trained from scratch on WIDER FACE) → 98-point WFLW landmark ConvNet → 478-point 3D mesh distilled from MediaPipe → 512-dim embedding (MobileFaceNet + ArcFace on MS1M, LFW 96.9%) → MiniFASNet anti-spoof ensemble.
All custom (except antispoof). Detector, landmarks, 3D mesh, recognition — trained from scratch. Only anti-spoof uses MinivisionAI's MiniFASNet under Apache 2.0. Every weight is AES-256-GCM encrypted on disk and decrypted in-browser via WebCrypto.