Two apps built around real catalog music (Spotify and YouTube), as opposed to my AI-generated library.
Diapason: discovery
You pit artists against each other in head-to-head battles, then react to scenes. The artists that survive become a "letter" of recommendations. Spotify checks every song and artist so nothing is made up, and YouTube supplies embeds.
Mixtape: memory
Pick a year and swipe through six moods, and you get a roughly 20-song tape from that era. It's meant to sound like what was actually on the radio, not a modern "best of".
The architecture
Both use a session-as-engine pattern. The web app writes a job request to disk, an AI session picks it up and writes back structured candidates, and the app's own engine.js applies the deterministic curation rules and validates every pick against Spotify. The model proposes; the code decides what's real and what makes the cut.
Two more experiments are in the same folder: PerfectLoops (finds seamless loop points in a song) and Bootleg (lines movie dialogue up over full-length songs).