after drake dropped iceman, one of the criticisms i saw most was his use of similies, and i wondered, how has drake's rhyme schemes changed over his career?

could ask an ai maybe, but idk how accurate it would be it cant just search all 550 drake songs and put them on a timeline.

so i figured id build it. so i started this idea "bythelyrics". the goal is to create an interactive website where you can ask questions about an artists music catalogue, and get answers including visualizations, that you can then export or share via a link

starting this thread just to document the development progress, if i dont share it somewhere ill give up.

streamable.com/n21fy1

heres where we're at today.

• have an ingestion pipeline using public apis to ingest an artists song catalogue and lyrics
• have an llm annotation pipeline that annotates all of an artists lyrics, marking entities (ie. places, people), sentiment (ie. moods), and more
• have a website where users can ask questions to an llm and it generates a query and a visualization against the available data
• users can pin these visualizations and export as an image to share on social media

next steps:

• a bit of polish, its still very data oriented, i havent done anything to make the wording nice, or to make the UI nice.
• automated ingestion to support more artists, currently i have to run ingestion once hourly for only 50 songs to stay under api limits
• improve quality of annotations, for example it classified "Controlla" as a person, obviously its not.
• improve the UI overall, idk im not a designer but ill figure it out

lotta fake songs on genius too

yeah, ill have to implement some type of cross reference. i think there is also going to have to be some type of manual intervention when llm confidence is low for other things. ill worry about that later. id rather get the core flows working then adjust data quality after.

truuuuu