What it answers

Spotify will tell you a song is 122 BPM in A minor with high danceability. It will not tell you that people put it on gym playlists and heartbreak playlists, which is the more interesting fact. A playlist is someone deciding a song belongs somewhere — do that a million times and you get a map of how music is actually used rather than how it is catalogued.

What it runs on

One million playlists, 3.6 million distinct tracks and 66.3 million playlist-track rows, with the Spotify Million Playlist Dataset as the spine and MusicBrainz, ListenBrainz, Last.fm and Deezer filling in identity and enrichment. Roughly a fifth of tracks carry a resolved ISRC or MusicBrainz ID, which lets release years come from first-release dates rather than whichever reissue a metadata provider happened to hold.

How it is built

The backend holds no dataset in memory: DuckDB queries Parquet artifacts directly out of Cloudflare R2 and streams results, with FastAPI in front.
Vector similarity lives in Upstash; one feature is LLM-assisted through Groq and falls back to deterministic keyword logic the moment the model is unavailable.
The frontend is React and Vite, with motion and d3 handling the moving parts.
Similarity search runs over the most-playlisted tracks, so an obscure query returns an honest empty result instead of a confident fabrication.
The "Your Listening" room parses a Spotify export entirely in the browser — nothing is uploaded, there is no account, and nothing is stored.

What makes it different

Every view ships an evidence contract: the exact metric, its source, its coverage, and what it does not mean. Artist ancestry states outright that it infers nothing about influence; ubiquity states that playlist reach is not listener count.
Coverage limits are disclosed rather than papered over. Views served from precomputed snapshots say so, and any page that could be misread as popularity says as much on its face.
Data vintage is stated instead of implied. The playlist corpus covers titles written between 2010 and 2017 and cannot be extended, so the language features are a record of how people named playlists then, not now; the editorial archive runs to 2026 and does keep moving.
The corpus measures placement, not listening. Only the "Your Listening" room measures listening, because only there does the data come from the person using it.

Scope and limits

The playlist corpus is fixed to 2010–2017 and cannot be extended — nothing open replaces a million user-written playlist names.
Playlist placement is not the same as listening, and reach is not a judgment of taste.
Only about a fifth of tracks carry a resolved ISRC or MusicBrainz ID.
Similarity coverage is limited to the most-playlisted tracks.