Separating music in the browser to make DJ mixes | Try it with your own track
How a browser DJ grew from automatic transitions to four-part music separation and two-deck mashups.
ALL POSTS
Practical AI experiments, architecture notes, and small tools you can use.
How a browser DJ grew from automatic transitions to four-part music separation and two-deck mashups.
With supervision and guardrails in code, the line still waited on my judgment. Its cost splits into waiting and attention, and risk decides where a person stays.
Rules written as prose couldn't stop the agents. What held was poka-yoke: judging an action by its effect and letting through only evidence that can be checked.
Why the AI manager missed things at volume, and how each station came to report its own stops.
Coding agents made commits grow tenfold. First I was the bottleneck; then the AI manager I built to replace me became one.
A 5.3MB Japanese voice clone, built from 6.5 minutes of my recordings, runs in a static page with no server. The experiments, diagrams, and the next limit.
Six San Francisco AI hackathons in two months, three wins, and what changed between them: a team of two, a tool with no API, and a live demo.
A full reset restores your allowance but can push the next weekly refresh back. A $46 model helps compare the timing.
Apple Speech and Translation meet Nemotron 3 speaker diarization in a local live-caption experiment.
Focus is not the same as covering the workflow a customer actually buys. A B2B product lesson about narrowing the business and spreading to the edges.
An AI mapped a product journey, captured the screenshots, and left the friction it found as cards on a shared canvas.
A visual guide to attention, the KV cache, and memory shared across layers.
Matched full-canvas results with 56% fewer tokens across 24 paired runs.
One open file, used in a custom app and Obsidian.
Make the last wording change yourself. Keep the editor in the file.
A reusable skill for slides you can share as HTML or print to PDF.