Current Method
I have tried for several times now to export various playlists, analysed them using ChatGPT or some other model (have tried a few models and a few generations), and then use that analysis to create new playlists that I use to discover new music. The analysis could contain things like mood, genres, artists, a certain dynamic range, etc. The results have been surprisingly good, even though some guidance is often needed, often depending on what model is being used. I then use soundiiz.com to transfer the playlist to Qobuz and then copying it to Roon.
Feature Description
I think one useful feature, at least for me, would be to be able to create new playlists based on entire playlists. I have previously used Roon Radio for this purpose, but it does seem to focus on the last song played and follow a quite generic script. I have tried to play a specific song several times to let Roon Radio suggest new songs, but the results are always the same. This suggests that the algoritm only takes the latests song into consideration.
One optional function could be that the user select what parts of the analysis of the playlist that should be taken into consideration when creating the new one. Also, the user could choose to focus the new playlist on discovery of new artists, exploring unknown songs from the same artists, etc.
Information and Features Unique to Roon
Roon have quite a lot of information on my history, what genres and artists that I have played recently and, perhaps the most important data point, access to my library, especially what I have added recently from streaming services or music that has been bought (I am old-fashioned and tend to want to own the music I really like), what music I have added to playlists recently, etc. This should make the generation of such playlists even more adapted to the tastes of the user than the single playlist method I have used with external AI:s.