← Work

Live · 2026-02 → ongoing

Ableton Mentor

AI augmentation in a non-traditional place — an MCP server that listens to my music sessions and tells me why track 7 sounds dark and mono.

Role: Solo build · Python / MCP / Ableton LOM / real-time DSP

The bet

AI augmentation isn’t just for code. A producer’s tools should reason about music, not just transport bytes. The same patterns work anywhere a domain expert wants software to think alongside them.

What it is

A daemon that traverses the Ableton Live Object Model, analyses audio in real time across three threads, and returns specific musical observations — not generic mix feedback.

Decisions that mattered

  • Three-thread architecture. LOM read, audio analysis, MCP server — separated so analysis never blocks the transport read.
  • Reference-track memory. Sessions remember the reference. New tracks evaluated against what the user said sounded right.
  • Specific over generic. Output is “EQ8 notch at 3.09 kHz” — not “the mid-range needs work.” The system either has the data or stays silent.

What’s next

A larger reference corpus. Maybe a way to teach the system new genres without retraining. Built for me; shared because it’s fun.