Our Edge-AI Lab takes a photo you choose, estimates a depth map, renders it as a 3D mesh, and lets CLIP describe the scene — and it does all of that without your image ever leaving the tab.
Why bother?
Privacy by construction. Nothing is uploaded, so there is nothing to secure, retain or explain. For regulated and consumer-sensitive products this is not a nicety; it is the difference between shippable and not.
Zero marginal cost. Cloud inference bills per request. A model running in the visitor’s browser bills nobody — the economics of a free tool like our labs (or a high-traffic feature) flip entirely.
Latency you don’t apologise for. No round trip, no queue. On a mid-range laptop the depth model runs interactively; on a phone it degrades gracefully.
The compatibility ladder
We target WebGPU first, WebNN where the platform exposes an accelerator, and fall back to WebAssembly everywhere else. The ladder matters more than the top rung: the fallback is what lets you promise the feature to every user rather than a hardware-lottery subset. The same discipline carries to our Android and embedded work, where NPU coverage is even patchier.
Try it yourself in the lab — then imagine it on your kiosk, set-top box or app.