Series Trailer: Music Streaming and Personalization, Built From Scratch

In the infantry we briefed before every mission so nobody was guessing what came next. This is the brief.

From now through January I'm posting a series on music streaming and personalization. Not slides about somebody else's system. My own. I built a streaming platform called Billboard, end to end: 5,344 analysed tracks, 18 models, and a live recommender that answers 95 of 100 requests in under 24 milliseconds. I'm going to walk through every piece of it.

Here's the idea in plain terms. A good recommender is the guy at the supply house who sees your truck pull in and already has your box of 3-inch screws on the counter. He didn't read your mind. He watched what you bought for six months. That's personalization. This series is how you build that guy out of code.

What it covers, in order:

1. Song to numbers: tempo, energy, pitch and key, measured from the raw audio
2. Model infrastructure: how models get trained, stored, served and retrained
3. Recommendation: from 5,000 candidates down to the 10 you see
4. Search and ranking: typing "that sad gym song" and getting the right one
5. Music intelligence: genre, mood, vocals, and naming a song from a few seconds of sound
6. Audio quality: from the studio master file to your earbuds

The schedule: weekdays are the streaming series. Weekends are data science filler episodes, because I like digging through numbers about countries, money and cities, and you might too.

What recruiters and engineers get from following: A real platform, not a tutorial. Every number read from the system on the day it posts. The misses, too, and the fix for each one. Code you can run. Comment CODE on any episode and I'll send the repo. A formal paper at the end.

About that paper. Slides 4 and 5 are the production scorecard: all 18 models, and the score each one has to hit to run in a real product. Genre detection at 82% against 42% for plain guessing. Song recognition at 99.5% on 2,000 clips. The recommender answering in 40 milliseconds or less. Those are the production bar, not today's results. When I get around to publishing the formal paper, it'll show the measured scores after every model is trained properly and production ready, graded against that table.

Who this is relevant to: engineers, data scientists and hiring teams at Spotify, Apple Music, YouTube Music, SoundCloud, Deezer, Netflix, Paramount+, Disney+, Hulu and Max. Music or video, it's the same job. Figure out what this person wants next, and get it there fast.

Part 1 is up next: Spotify took away danceability, energy and tempo. I rebuilt them from raw audio.