So I built Atomic Afterglow. It’s a Python-based analyzer that runs in the browser but processes audio locally (zero retention/no uploads).

How it works:

Engine: Built on Librosa and Streamlit.

Drop Detection: Instead of finding the "quietest" section, it uses a rolling RMS window to find the deepest structural energy trough in the middle 60% of the track (the breakdown/mix point).

Harmonics: Uses the Krumhansl-Schmuckler algorithm to generate the Top 3 Key candidates (Camelot formatted) with confidence margins, rather than forcing a single result on ambiguous tracks.

Repo (Open Source): https://github.com/phasesequencer/atomic-afterglow

I'd love to hear if the key detection holds up on your more complex or detuned tracks.