← How I AI

Replaced a four-app DJ prep chain with one step, and AI plans the night

Built with
  • OpenAI
  • Claude Code
  • SwiftUI
  • JavaScript
  • Python
  • MusicKit
Status
Testing

The party is tonight. The playlist isn’t made yet.

  1. Labels every song
  2. Levels the volume
  3. Finds what mixes
  4. Builds the playlist
Sources
My files, Apple Music, Spotify, YouTube Music
Sends to
djay Pro, Serato DJ, rekordbox, Traktor, VirtualDJ

The short version

Context

Rip is a Mac app I’ve been testing on my own DJ prep, and it’s working out well so far. It keeps every music source in one library, and its AI curates the night’s playlist for my DJ software. After a lot more testing, I plan to publish it on the App Store.

What I did

On my own, with Claude as the pair programmer, I’ve designed and built:

  • The Mac app, its web library and an iPhone app
  • The audio engine, and my own energy model
  • The AI that labels, searches and plans a night

What’s shipped

In testing on my own Mac and iPhone, with my own library of about 1,000 songs.

Impact

For my own house parties

Hours
to curate one party’s playlist by hand
Seconds
with Rip, from one sentence

Problem

Every party I DJ started with hours of curating. Now it starts with one sentence.

Too many tools, and too many files moving between them. Platinum Notes levelled the volume, Mixed In Key found the key and BPM, I sorted the folders by hand, and then I built the playlist song by song.

What I did

I designed one app to replace that chain, and built it with Claude as the pair programmer. I trained a small model on about 830 of my own files to score energy, and designed how the AI labels a song, reads a search and plans a night. Where the AI isn’t sure, it asks me.

The AI features I built

Planning a night

A playlist from one sentence

I describe the night and Rip builds it, in order, in seconds.

Search in plain words

A vibe, a language and a tempo, typed the way I’d say them.

What mixes well

Open a song to see the ones that fit its key and tempo.

Knowing every song

Labels on every song

The AI tags each song by vibe, language and era.

Energy from my own model

A model I trained on about 830 of my files scores each song.

When the AI isn’t sure, I pick

Two answers side by side, and my answer always wins.

Tech stack

AI

  • OpenAI
  • Claude Code

Built with

  • Claude Code
  • Xcode

Front end

  • SwiftUI
  • JavaScript

Engine

  • Python
  • Essentia
  • Mutagen

APIs

  • OpenAI
  • MusicKit
  • iTunes Search

Evals

Testing Rip against the tools it replaced

Before I trusted Rip with a party, I ran the same ten songs through my old tools and through Rip, then compared what each one wrote into the file.

Model results
MeasuredMatched my old toolsDiffered
Tempo (BPM)10/100
Key8/102

What I found

Tempo matched on all ten songs and key on eight. Of the two keys that differed, one was the major of the same Camelot number, which still mixes.

What’s next

Where it stands

I’m testing Rip on my own DJ prep, and it’s working out well so far. After a lot more testing, I plan to publish it on the App Store. Next, the iPhone app catches up with the web app.