02AI · Music
AutoMix
Solo build · 2025
Automatic beat-matched mixing, driven by a learned model
17blend / transition strategies
Blends tracks into one continuous mix, using a dual-stage model to pick the technique and generate the automation curve for each transition.
A dual-stage architecture first picks the mixing technique, then generates the automation curves across 17 blend and transition strategies. A separate 3-stage data-analysis pipeline handles QA, testing the transition logic against published work on mixing rather than hand-tuned rules.
Stack
- PyTorch
- LSTM
- scikit-learn
- Librosa
- Demucs
- NumPy
- Pandas
02
How it's built
SPECIFICATION
- Model
- dual-stage: technique + automation curves
- Strategies
- 17 blend / transition types
- QA
- 3-stage data-analysis pipeline
- Audio
- Librosa · Demucs stem separation
MY ROLE
- 01Built the dual-stage model across 17 blend/transition strategies.
- 02Generated per-transition automation curves from the audio.
- 03Architected a 3-stage QA pipeline validating the transition logic.