Teaching a machine to build a DJ set

This site is organised like a record: side A, tracklists, a sleeve for every project. That is not only a design conceit. Sequencing a DJ set is the same problem as sequencing an interface — you are deciding what the next thing has to do to the person on the receiving end, and you are constrained by what just happened.

AIDeeJay analyses a local music folder and builds sets from it. Essentia does the signal work: BPM, key and scale, Camelot notation, loudness and dynamics, spectral and timbral features. A text-only model pass then labels each track with mood, relative energy from low to peak, and a role in a set — opener, builder, peak-tool, bridge, cooldown.

Harmonic mixing is treated as important but not sacred. A non-harmonic transition is allowed when BPM, energy, and role justify it, because that is how the good ones actually work.

Local-first, on purpose

No audio is uploaded anywhere. The analysis runs on the machine holding the library, and the output is a plain library.jsonl you can read, diff, and regenerate. There is now a SwiftUI companion that runs the same pipeline on device — scan a folder, analyse, build a set, export the library, the set, and an m3u8 playlist to app storage.

What it is really for

The interesting part is not automating a set. It is that the analysis makes explicit what a DJ carries implicitly: this track is a bridge, that one is a peak tool, this pair only works because the energy is still climbing. Making a tacit judgment machine-readable is the same instinct that makes me pull apart a renderer — the goal is not the machine, it is the vocabulary you get afterwards.

Both repositories are currently private.