I'm a fan of three things: electronic music, DJing with samples, and Web AI. This talk is about what happened when the third solved a problem I had with the first two.

The sample you want is almost always locked inside a finished record: a bassline buried beneath a vocal, four bars of piano that were never released on their own. For years, the answer was an EQ, a filter, and a lot of hope.

Then source separation models got good, ONNX runtimes got fast, and WebGPU turned a laptop into a machine that could actually run them. So I built an app: drop in a track, get its stems (bass, drums, vocals, guitar, synth), then cut out the four bars you actually wanted. It runs entirely in the browser, and nothing is ever uploaded.

In 15 minutes, I'll explain how it works, then tell you:
- why the mainstream Web AI library couldn't host this model
- what cross-origin isolation gives you, and what it quietly breaks
- an assumption I was convinced was true, until measurements proved me wrong
- the app URL, so you can have your own audio studio in a browser tab

A couple of years ago, this would have required a data center. Now it takes a laptop and a browser tab. That gap is the story of Web AI.