about

About xz.

I'm a developer in Tokyo. I make small things where software and sound meet, and tools that let AI models do real work on my own machine.

What I make

Most of what I build starts as a question I can't stop poking at. Can a language model write music that sounds like one song rather than a string of ideas? Can a 27-billion-parameter model run fast enough on a Mac to be useful every day? Can a Mac keep a queue of jobs for local models and hosted coding agents running through restarts, without me babysitting it?

The answers turn into projects. Two of them are on the home page:

Why I write here

Building things with language models is still mostly folklore. Half of what I learn is a small, specific trick — ask for the rhythm as blank slots instead of in words; generate three options from three different briefs instead of one prompt three times; put the model proxy on the same origin so iOS Safari will stream it. Those tricks are easy to forget and hard to find, so I write them down in Writing, with the real numbers from my own tests where I have them.

Everything on this site is made by me. The audio on the home page is real Tunepick output, rendered from the same arrangement that draws the piano roll, not a mock-up.

How this site works

zxa233.com is a static site served by a small Cloudflare Worker. The only moving part is the melody guestbook on the home page: you can leave an eight-note tune, no words, and it shows up for the next visitor. The fonts are self-hosted, and there is no analytics script. The site shows ads through Google AdSense; the privacy page explains what that means for you.

Say hello

The quickest way to reach me is on GitHub. More options are on the contact page.