Also, other comments have suggested that this was trained on 120MM of the audio previews instead of the full songs. That might explain why the recommendations seem a little off for some people.
I tried two of my all-time favorite songs:
The Gaslight Anthem - Handwritten
Thrice - The Artist in the Ambulance
On the first, I would say it was a total swing and miss - the recommendations had similar elements to Handwritten like strummed power chords, but the vibe was completely off. With Thrice, same thing. The first match was a remaster of the song itself by the band, so of course despite being a "different" track it was the closest possible match. The rest of the recommendations had a similar tempo and more heavy metal riffs (as are common in Thrice's songs), but none of them were songs I would voluntarily listen to.
This is a cool idea and I hope OP if you read this you take the criticsm as it's intended - I'm not trashing the implementation or anything. This just feels like it truly is a robot that can't feel the emotion of a song making me recommendations.
The current model I have isn't as good as I want it to be, and I'm working on a newer one with a different training process as well. This should address some of the shortcomings people have mentioned.
I was really nervous of shipping my v1 with the current model, but thought I might as well share what I have so far with the world, in case someone finds it useful lol.