Auxology

Context for LLMs

Auxology is a personal music-taste engine and playlist product, designated GSV-B4 (pace-layer B, life-systems domain) in the General Systems Ventures registry. It joins three inputs, a person's Last.fm scrobble history, the EveryNoise genre graph, and an expert-feedback loop, and its key design decision is treating each playlist as a situated, per-venue profile rather than a single global model of one person's taste: the founding use case is running the aux at a specific room, for a specific owner, with a specific function, so playlists are built for a context, not for a person in the abstract. The pipeline begins with data acquisition: export.py, aggregate.py, and auth.py pull and summarize full Last.fm scrobble history, while enao_scrape.py scrapes EveryNoise At Once for 6,258 genres and roughly 457,000 artists with acoustic coordinates, giving the system a large structured map of genre-space to place a listener's taste within. build_embeddings.py and cluster_modes.py derive personal artist embeddings from a word2vec co-listening space and partition them into k-means taste modes, while disambiguate.py resolves artist-name collisions before they corrupt the signal, a data-quality step that matters because the embedding and clustering stages depend on clean artist identity. Rather than training once, the taste model operates as a continuous loop: playlist_recs.py exposes init, recommend, feedback, refit, and status verbs per playlist, and verdicts are captured through triage_app.py and a triage.html web UI with embedded YouTube playback and reason-tag chips usable by touch or keyboard. Refits use logistic regression once a profile has at least fifteen labels, falling back to Rocchio below that threshold, and status reports hit-rate. It sits in the life-systems-heart and media-ecology strands of the GSV portfolio, alongside related projects such as underground-beets, vibecology, and Horza. Status: in-flight as of 2026-06-11, with a substantial working pipeline (export, scraping, embeddings, clustering, feedback loop) already built and documented.