This is toposonico, a music recommender and navigable map. At core it's a skipgram word2vec model trained over ~6M playlists. Tracks are embedded in a 128d space. Embeddings for albums, artists and labels are computed marginalizing over tracks. The 2D map was built with UMAP.
Both the model and UMAP were trained in the cloud over a NVIDIA A100. All things considered it cost me around ~50EUR, over two main training sessions and a few experiments. For the slippy map I experimented with a few libraries. Ended up with Maplibre GL JS. Loved working with it, kudos to their developers. For the recommender indexes I used FAISS, another great piece of software. Pretty happy with the thing running on a small and cheap box. It was the first project where I used LLMs extensively. Still I wouldn’t say this was vibecoded: getting the thing to work as I wanted was the usual pain in the arse and I ended up rewriting by hand multiple times. Mostly Claude on Pro plan when the plan was still usable, then I switched to codex (which eventuay became unusable too, but that’s a another story).
It still has some problem. For example very big artists often ends up very disconnected from the rest of the embeddings. I guess this has to do with the high number of mono-artist playlists these ends up on. But I’m not very sure still, and would like to know if anyone else faced a similar issue and how it was fixed. I’m interested in any kind of feedback on UI/UX. Love working on these but, coming from a scientific computation background, have little experience.
Repo link: https://github.com/peppedilillo/toposonico
No comments yet.
