Spotify recommended me 34 separate murder podcasts after I listened to one Crime Junkie episode months ago, and I took that personally. Instead of skipping like a normal person, I decided to build my own audiobook recommendation system that’s based on me, not whatever engagement metric is trying to turn me into a detective. I started with audiobooks because they’re long and high-stakes—recommend the wrong one and you’ve wasted 12 hours of someone’s life. I pulled a million+ books from the Goodreads Kaggle dataset, embedded titles + descriptions, and built a FAISS vector search engine to find similar books to my actual taste. My first run was a disaster (13/15 fantasy recs… I don’t read fantasy), and the bug was painfully simple: my “books I like” only had titles, so the search panicked and returned dragons and Greek tragedies. I fixed it by fetching missing descriptions via the Google Books API, then deployed the whole Python script to Modal so it could run as a serverless endpoint. Finally, I used a Notion custom AI agent (my life lives in Notion) to trigger it at 6:30am, send 15 candidates to Claude, verify availability with Spotify’s API, and push the daily pick to my phone. Overkill? Absolutely. Regret? Not even a little.