Browser-based hand gesture T9 keyboard (YOLOX and ONNX Runtime Web)

Name: Browser-based hand gesture T9 keyboard (YOLOX and ONNX Runtime Web)
Availability: InStock
Author: huang4fun

by huang4fun·Feb 18, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidWizardryNiche Gem

The Take

Runs everything in the browser and actually stays responsive — ONNX Runtime Web + a YOLOX model handle subtle hand-seal distinctions that MediaPipe struggled with. Clever choice to layer a T9 keypad over gesture input (reduces required class count and makes errors tolerable), but the demo remains an experiment: lighting sensitivity and similar seals create real UX friction and it’s not yet a drop-in input alternative.

Post Description

I built a small experiment over a 3-hour vibe coding session: a real-time T9 keyboard controlled by hand gestures, running entirely in the browser.

It uses:

YOLOX for gesture detection

ONNX Runtime Web for in-browser inference

Plain JS for the UI

The original goal was simple: Could I make real-time gesture-based input usable inside a browser without freezing the UI?

A few observations:

In-browser ML performance is better than I expected on modern laptops

Subtle gesture distinctions (e.g. similar seals like Tiger vs Ram) require stronger detection than MediaPipe provided — YOLOX performed noticeably better

Lighting consistency matters more than hand size

It’s obviously not production-grade, but it was an interesting exploration of browser-based vision input.

Curious what others think about gesture interfaces as alternative input systems.

Demo: https://ketsuin.clothpath.com/