No audio dependency
The classifier operates on metadata and tags, so it can run without storing copyrighted audio files.
AI-GENERATED PROJECT / 2026
Spotify Radio Architect
A Python classification and publishing tool for a 2,604-track Spotify playlist.
This write-up—was generated with AI.
2,604tracks classified
20destination playlists
233held for review
0remote mismatches
01 / CONSTRAINTS
What the system had to work around
Spotify supplied playlist reads and writes, but this development-mode app did not have dependable per-track genre or audio-feature data. Spotify therefore became the source and destination layer, while Last.fm supplied community tags for tracks, albums, and artists. The tool does not download or inspect audio.
Last.fm tags are noisy
Coverage differs by song, counts are user-generated, and some tags describe fandoms rather than music.
The source is immutable
The source playlist ID is rejected as a publish destination. Every write goes to a generated playlist.
Network calls need state
Spotify pagination, 100-item write batches, Last.fm rate limits, and retries are handled explicitly.
02 / ARCHITECTURE
A local pipeline with inspectable artifacts
01 / EXPORT
Spotify snapshot
OAuth PKCE, paginated reads, normalized track records, and a saved source snapshot.
02 / ENRICH
Last.fm cache
Track, album, primary-artist, and secondary-artist tags are cached separately.
03 / CLASSIFY
Rule engine
Weighted evidence is scored against 19 programmed stations and a review bucket.
04 / REVIEW + WRITE
Report and publisher
HTML and CSV inspection precede dry-run, batched writes, saved IDs, and remote verification.
data/source.jsonstores the normalized Spotify export.data/lastfm-cache.jsonprevents repeated enrichment calls.output/assignments.jsonstores tags, every station score, the winner, and confidence.data/published.jsonmaps station IDs to destination playlist IDs for repeatable updates.
03 / CLASSIFIER
Track evidence is allowed to overrule the artist
The initial version treated artist genres as the main signal. That failed on artists whose catalog spans several styles. Version 4 uses an evidence hierarchy and changes the multipliers based on what is available for each individual song.
1.00 track / 0.58 album / 0.14 artist
The track dominates. Album and artist tags only add context.
0.84 album / 0.24 artist
The album becomes the main signal when the track has no useful tags.
0.72 primary artist
Artist evidence remains usable but carries less authority. Featured artists receive another 0.45 multiplier.
Strongest vote plus limited reinforcement
Duplicate evidence keeps its strongest weight and adds 18% of the weaker supporting vote.
Each station defines primary tags, secondary includes, exclusions, optional year ranges, and whether
a primary tag is mandatory. Generic labels such as pop, rock,
hip hop, and electronic receive lower multipliers, especially when they
come from artist-level evidence. Exclusion weight reduces a score multiplicatively, and final scores
use log1p to limit the effect of many overlapping tags.
A song is assigned only when the winning score reaches 0.19. Confidence is recorded as
(winner - runner_up) / winner. The current configuration keeps one destination per track,
but stores every score so overlap can be reviewed later.
04 / EVIDENCE COVERAGE
How the 2,604 assignments were decided
Track tags1,001 / 38.4%
Album tags with artist context732 / 28.1%
Artist fallback638 / 24.5%
Missing or below threshold233 / 8.9%
The review report exposes the selected station, confidence margin, evidence source, normalized tags, and all candidate scores. A CSV override file can replace any assignment without changing classifier code.
05 / OUTPUT DISTRIBUTION
The final station set
Mainstream pop, dance-pop, and radio-sized hooks.
Pop-rock, adult contemporary, acoustic pop, and songwriters.
Indie, art, bedroom, and alternative pop.
K-pop, J-pop, C-pop, and East Asian hip-hop.
Synthpop, new wave, electroclash, and retro-futurism.
Hip-hop, boom bap, conscious, underground, and classic rap.
Trap, drill, cloud rap, emo rap, and southern rap.
Contemporary R&B, alternative R&B, neo-soul, and slow jams.
Alternative, indie, garage, grunge, and post-grunge rock.
Punk, pop-punk, emo, skate punk, and post-hardcore.
Metal, metalcore, nu metal, industrial, and hardcore.
House, techno, trance, progressive house, and club music.
Dubstep, drum and bass, future bass, and breakbeat.
Ambient, downtempo, trip-hop, chillout, and shoegaze.
Soul, funk, disco, Motown, boogie, and warm grooves.
Country, folk, Americana, roots, and storytelling.
Reggae, Latin, Afrobeats, dancehall, ska, and global music.
Classic, blues, southern, psychedelic, and arena rock.
Soundtracks, orchestral scores, classical, and instrumentals.
Missing, ambiguous, or below-threshold evidence.
06 / TRADEOFFS
Decisions that affected the result
Deterministic metadata won
It is reproducible, cheap to rerun, and easy to inspect. It cannot detect tempo, timbre, or mood directly.
Uncertain tracks stay uncertain
Required primary tags and the 0.19 floor reduce false positives but leave 233 tracks for review.
Enrichment is cached
The cache contains 2,566 track, 1,986 album, and 1,609 artist records. Refreshing is explicit.
Assignments are exclusive
This avoids duplicate playlists at the cost of hiding some genuine cross-genre relationships.
The last taxonomy change split a broad pop station into Headliners and The Avenue. Removing
pop rock and adult contemporary from Headliners produced a separate
127-track station without introducing overlaps.
07 / PUBLISHING
Writes are explicit and repeatable
- Authorization Code with PKCE is used, so no Spotify client secret is stored.
analyzeandprevieware read-only; onlypublishwrites.publish --dry-runlists every intended create or update.- Destination IDs are persisted, so another publish updates the same playlists.
- Track URIs are written in batches of 100 with retry handling.
- A read-only verification pass checked all 20 remote totals: 2,604 tracks and zero mismatches.