Percept
Image recognition that runs entirely in your browser.
Overview
Percept is an AI demo that runs a real trained neural network — MobileNet — directly in the browser. Upload any image and it predicts what's in it, with confidence scores, in about a second.
Because inference happens on the device, no image is ever uploaded to a server. It's a genuine machine-learning showcase, not a mockup — the model recognises 1,000 object classes.
What it does
Screenshots
Why we built it
AI features are often demoed as mock-ups that don't actually do anything. Percept is the opposite — it runs a real, trained neural network (MobileNet) directly in your browser and tells you what's in any image, with confidence scores, in about a second.
The purpose is to prove genuine machine-learning capability and a privacy-first approach: because inference happens on-device, no image is ever uploaded. It's the pattern we'd reuse for any client feature where data must never leave the user's device.
How we built it
Pick a real model
We chose MobileNet — a genuine, production-grade vision model — over a scripted fake.
Run it on-device
Loaded with TensorFlow.js so inference happens in the browser on the GPU, with zero uploads.
Design the results
Top-5 predictions with animated confidence bars that make the model's reasoning legible.
Ship it live
A drag-and-drop demo anyone can try with their own images.