originally a school project for implementing shared queues in spotify, turned into a knowledge-graph visualization of how much your music taste overlaps with your friends'.
you log in with spotify, your friends log in with spotify, and the app draws a graph where distance means dissimilarity — the closer two people sit, the more their listening actually lines up. the edge label is an honest, calibrated "% match", and you can click an edge to see why you matched.
unfortunately, the app is under developer mode in the spotify developer platform, and as of april 2025 i cannot qualify to request extended quota/access
the first version was rough, and honestly a bit dishonest: taste was
0.5·artist-jaccard + 0.5·genre-cosine, and the lyric score was the cosine of the averaged
embedding over your ~20 most-recent lyrics. averaging anisotropic embeddings collapses every
pair to ~0.85–0.99 cosine — no dynamic range, so the number meant nothing. binary artist
overlap also scored literal bandmates as 0.
v2 represents a person as a rich, normalized, distributional object built from what they actually play over time, not a top-N snapshot:
- continuous footprint — every play decays with a 30-day half-life, so the profile is a recency-weighted picture of your listening, not whatever spotify surfaces as your "top".
- whitened embeddings — we remove the common anisotropic direction so cosine is meaningful again. this is the whole reason the lyric facet works now.
- set / distribution distance — lyric themes are compared with word mover's distance (entropic sinkhorn / EMD) over per-track embeddings. never averaged.
- bayesian facet blend — artist / genre / lyric facets are blended with data-derived
weights (
softmax(-τ·loss)), and the per-facet contribution is returned with the score. the blend is the explanation. - calibration — raw scores map to percentiles over the population of pairs, so "73% match" is honest.
the spine is the content × behavior bridge — lyric-theme content crossed with listening-history behavior — which the content-similarity survey literature flags as barely explored. it's the project's defensible angle. all of it is grounded in real work; see docs/EVIDENCE.md.
the app UI stays intentionally abstracted (it's a vibe, not a dashboard). these figures —
generated by running the real engine on synthetic listening data
(python3 scripts/make_figures.py) — are where we show what's underneath.
the bug, and the fix. averaged embeddings pile up near cosine 1.0 (no range); whitening restores the spread, so similarity becomes informative:
recency. every play decays with a 30-day half-life — continuous, not a snapshot:
word mover's distance. the lyric facet aligns one listener's tracks onto another's (the optimal-transport coupling), instead of averaging both into mush:
the blend is the explanation. every match decomposes into per-facet contributions — this is the comparative-imagery panel you get when you click an edge:
the graph itself = whitened profiles projected to 2D; distance ≈ dissimilarity:
every visual channel maps to a real quantity: position = taste, distance = dissimilarity, edge label = calibrated percentile. nothing decorative pretends to be data. the full design language (and the aesthetic — mono / ASCII, "claude code" terminal vibe) lives in docs/DESIGN.md.
- engine: pure-stdlib python taste engine (
src/common/taste/) — lambda-ready and WASM-portable; heavy fitting (whitening eigendecomposition, calibration quantiles, ensemble weights) is offline numpy intaste.fit. 52 unit tests, incl. a dynamic-range guard that fails the build if whitening ever stops working. - frontend: react + typescript + d3, hosted on s3 / cloudfront
- backend: python lambda behind api gateway (aws sam)
- storage: dynamodb for profiles, play history, and friend graphs
# unit tests (stdlib only)
python3 -m unittest discover -s tests/unit -v
# regenerate the figures above (stdlib + numpy, no matplotlib)
python3 scripts/make_figures.py- a rust → wasm kernel that runs the pairwise scoring + 2D projection client-side (the one deliberate native-code showcase)
- the V1/V2/V3 interpretability UI from docs/DESIGN.md
- parked behind clean seams: an acoustic "waves" facet (spotify audio-features went dark post-2024) and population-scale triplet metric-learning (waits on user volume)
contributions and suggestions welcome — especially richer embeddings / similarity math, and anything that makes the graph more interactive and more interpretable.
Rodrigo Sastré