Personal project · October 2026
Sleeper
Pick five albums you like and get one niche album to play next, plus four runners-up.
- Stack
Problem
Most recommenders are either inaccurate or keep pointing you at the most popular picks. Sleeper is built to find “sleeper picks”: pick 5 albums you like and get 1 niche album to play next plus 4 runners-up, each with a one-line reason built from a scoring index.
Approach
The engine is a pipeline of small, testable steps:
- Resolve each pick to a MusicBrainz release group, so deluxe editions and remasters are the same album.
- Profile the picks from cleaned Last.fm tags, with junk tags like “seen live” removed.
- Cluster mixed picks (two rap, two jazz, one shoegaze) and recommend for the biggest group instead of averaging into mush.
- Score candidates by tag similarity plus similar-artist overlap, minus a popularity penalty.
- Explain with templates filled from the signals that produced the score. There’s no LLM, so the reason always matches the ranking.
Every API call goes through a SQLite cache and a per-API rate limiter, and picks live in the URL so results are shareable without accounts.
Result
It’s live, with CI and a pytest suite. A leave-one-out evaluation gives a 45% album hit rate in the top 20, against 47% for a naive similar-artists baseline. That gap is the point: the evaluation rewards predicting albums you already like, and the popularity penalty deliberately trades a little of that for picks you don’t know yet.