Building in the open

Dev Journal

Notes from building MoVox — the choices, the engineering, and what’s next.

• November 5, 2026 A series · 3 parts so far

  After chapter one

  Chapter one went out as a test. These are the lessons from what came after: listening to the people who watch, deciding what's worth doing, and building the checks that let one person keep up.

  1. 一 Ship it to real people Chapter one went out as a test, and the people who watched it changed the product three times: a version of the story they could follow, pictures that are actually real, and episodes instead of chapters. Here's what building around the people who use the thing looked like.
  2. 二 One list for everything The 白话 version of MoVox was a one-to-two-week job, text and audio only. It took four, and it came back with photoreal clips and lip-sync. Here's the sprint rule I've taught teams for years, and how I caught myself running it on a project of one.
  3. 三 Stop being the check For two chapters I reviewed every picture by eye, and every flag got fixed by hand. By the end there were 136 hand fixes, and the next chapter would need them all again. Here's why a fix that never reaches the process gets made twice, and what moving it there actually took.

• October 27, 2026

  Read the license before you watch the demo

  I promised this post three times. Every model in MoVox is Apache 2.0 or MIT, and the reason is practical, not idealism: I want the finished film to be mine to publish. Here's what that rule cost me, what it ruled out, and why the license on the repo is often not the license on the weights.

• September 23, 2026 A series · 4 parts

  Two months quiet

  I went dark on a build-in-public project to make a demo. Four notes on what happened instead: the pictures, the book, the voice, and a second computer.

  1. 一 I went quiet to make a demo I vanished for two months because I wanted a clip I could post. The pipeline can draw now, and chapter one plays end to end. The pictures still don't pass my review, and that's the honest status of the demo.
  2. 二 Sixteen thousand characters gone In June an audit found 15,753 characters missing from the book. I fixed it, got to 99.96%, and told you in July that the Reader works. This summer I found out 99.96% is not the book. A contract the model won't keep has to be enforced in code.
  3. 三 Six minutes is not a voice I wanted the narrator to be me. A real reference clip already sounds like me. Fine-tuning on six minutes of isolated sentences taught the model my timbre and then made me shout. I stopped tuning.
  4. 四 It's two computers now In July I said MoVox ran on one machine on my desk. Now it runs on two. The second one is a 32-gigabyte MacBook, and it didn't buy me better pictures. It bought me a second card, which is a different thing.

• July 25, 2026

  A text file goes in, a film comes out

  What MoVox actually is, end to end. Five AI agents, one human signing off at every checkpoint, all of it running on one machine on my desk — plus an honest account of which parts work today and which parts are still a gray rectangle.

• July 23, 2026

  The book was too heavy

  How my love of the Chinese classics began — a Journey to the West too heavy for a second-grader — and why, decades later, I'm teaching a computer to read those stories to my kids in my own voice.

• July 23, 2026

  AI is just the next tool

  Two halves of one stance: AI isn't coming for your job, and what it makes isn't slop. An engineer on why the next tool is just that — a tool — and how craft survives, drawn from building MoVox.