Qalarc / Answers

How do I run face recognition and face AI in the browser with no uploads?

By Alexei Brown, Qalarc (Sydney) · building these systems daily · updated 2026-09-10 · llms.txt
MediaPipe FaceMesh (468 landmarks, WASM) + ONNX Runtime Web runs real-time face AI fully client-side — recognition, expressions, demographics, even heart rate from ambient light (rPPG). No frame ever leaves the device. We run this stack in production demos with 14 self-trained models.

The stack that works in-browser

JobToolNotes
Tracking / landmarksMediaPipe FaceMesh468 points, 30fps on laptops, WASM
Model inferenceONNX Runtime WebWASM; WebGL/WebGPU where available
RecognitionEmbedding models + local storePersistent identity library, all local
DemographicsSelf-trained small modelsWe run 14 ("QalArc Folkus") — age bracket, gender, ethnicity, skin-tone zone
Heart rate (rPPG)Signal filtering over face ROIGreen channel carries the pulse

Verify it yourself (the network tab test)

Every demo below opens your camera — then shows no video frames uploaded — after the one-time model/runtime loads, the network tab stays quiet during use. That's the honest on-device proof; demand it of any "private" face AI you evaluate.

Face tracking demo Heart-rate / rPPG demo The platform writeup

What it costs

Model downloads of a few MB happen once; after that, inference is free and offline. CPU-only machines handle tracking + small models; GPU helps for multi-face plus the demographic stack simultaneously.

The lines that matter

On-device does not mean consequence-free. Consent for analysis, signage if scanning others in public, and honesty about accuracy limits — demographic models carry bias you must measure and state; we publish ours as an experiment, not a product claim. Never pipeline frames to a server while marketing "local".

When NOT to build this

Forensic-grade recognition or heavy multi-camera analytics still belong on server GPUs. The browser stack is for privacy-first, zero-install, single-user use — exactly what wearables and personal tools need.