About and pressquotable facts, no invented ones
Mixterio started as a personal tool, not a startup. A DJ in Georgia with a large library and a recurring problem: the software knew the filenames and the waveforms, and nothing about the music.
The library tools could tell him a BPM and a key, sometimes correctly. What none of them could tell him was whether these two tracks actually work together, whether this join would land, or which forty tracks in a folder of thousands belonged in one night. That knowledge lived in his head and nowhere else, and it did not scale past the records he had played recently.
So he built the thing that measures it. Analysis first, because you cannot reason about music you have not read. Then the part nobody else does: render the transition as real audio in advance and grade it, so a pairing is not a prediction but a measurement. Then the engine that performs only what passed, or exports it with transition cue points to Rekordbox and CDJs, at one locked tempo.
He is not a developer. The engineering was done with AI as the instrument: writing the code, running the measurements, and saying plainly when an idea failed. A good number of the ideas in this project are ideas that got measured and thrown away, and the honest results are on the board next to the ones that worked.
The engine has played a real gig, in a real room, for real hours - planned in the app, exported with cue points, and played live on CDJs. That is the standard the rest of it is being held to: not a demo that survives two minutes, but a night that survives a spilled drink.
Press facts
Mixterio is a desktop DJ engine that analyses your own library, composes full sets you can read before you hear them, and performs them with every transition rendered and machine-graded before it plays. It is the only DJ product that publishes reproducible, held-out analysis accuracy.
Other tools estimate whether a pair will work from key, tempo and energy. Mixterio renders the transition as audio and measures it, 22k+ pair configurations so far, and refuses to play the ones that fail.
72.67 Rekordbox 7measured by us, same 566 tracks
held-out, shipping app89.1 Beat This!Foscarin, ISMIR 2024
held-out, never trained on78.3 Beat This!Foscarin, ISMIR 2024
held-out, never trained onno published comparisonnobody publishes a figure for this metric, so we draw none
raw grid, general material0.472 Foote (2000)measured by us, same corpus
held-out, coarse HR3FHeld-out figures on public datasets. Full board, protocols and artifacts.
If you find a weakness we have not published, that is a bug in our honesty and we want to hear about it.
Press kit
We are not shipping mock-ups of an app you cannot install yet, or stock photographs of a DJ who has never used this. Screenshots go up when they are screenshots of the real thing, and demo audio goes up when the music in it is cleared.
In the meantime, everything factual about the product is on this page and on the board, and it can be quoted directly.