Back to Projects
experiment + web

On-Device AI Avatar

Tech Stack

TypeScript Vue.js WebML WebGPU Kokoro.js SpeechRecognition API WebGL (Three.js) VRM (Virtual Reality Model)

Links

Year

2025

The Challenge

Creating interactive AI-driven avatars typically requires server-side processing, raising privacy concerns and latency issues for users.

The Solution

A virtual avatar and AI prototype running completely local in the browser, leveraging WebAssembly and on-device machine learning models to ensure user privacy while delivering interactive experiences.

Technical details

Prototype of a Typescript/Javascript-based web application running entirely in the browser. (WebGPU)

  • Language: Typescript/Javascript
  • Framework: Vue3
  • Local LLM: Qwen, Gemma, Phi, Mistral etc (ONNX, quantized) running with WebML and WebGPU on Web Workers
  • Avatar: VRM (Virtual Reality Model) rendered with Three.js
  • Speech-to-Text: SpeechRecognition API
  • Text-to-Speech: Kokoro.js
  • Deployment: Vercel
?

Interested in this stack?

Let's build something similar.

Contact
Next Project