The usual pitch for on-device AI is speed — no network round trip. The pitch that matters more, especially for anything touching biometric data, is that data which never leaves the device can never leak from a server.
The architecture decision, concretely
In FaceVision, detection, embedding and liveness checks run entirely client-side via ONNX Runtime Web. The backend receives a 512-dimension embedding, never an image. If that backend is ever breached, there are no faces to steal — only vectors that can match, not reconstruct.
What you give up
Model size is constrained by what a browser can reasonably load and run, and you lose the ability to centrally update inference logic without a client update. For anything privacy-sensitive, that tradeoff is usually worth it.
Where this pattern applies beyond faces
Any input a user would reasonably not want stored — voice, handwriting, location patterns — is a candidate for on-device inference producing a derived, non-reversible representation instead of storing the raw input.
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU Kanpur) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. Email pranjulrathour41@gmail.com.
