Skip to content
Thoughtful Engineering
All work

Kanha

Website chatbots that run on your visitor's device.

Kanha crawls the pages you choose, trains a compact model on them and delivers it to the browser, so answers cost nothing per question.

The Kanha landing page with a peacock feather

Most chatbots call a hosted model for every answer. Kanha fine-tunes a small model on your own content and runs it with WebGPU in the visitor's browser. You pay for the pages you index and the training runs you start, not for each reply.

Platforms
Web, npm
Built with
Rust, Next.js, WebGPU, WebLLM
Inference
On device, with WebGPU
Embed
Web Component or React

The thinking.

  • First thought

    Call a hosted model for every answer.

    Second thought

    Train a small model on your pages and run it in the visitor's browser.

NextJambudvipaA world that remembers what you do.