Upcoming Talk

Walk Up Song

May 21, 2026Tech Lancaster at West Art

Innovation and creativity in the age of AI coding — told through the story of building something fun.

The Idea

It started at Five Guys on a Wednesday night. Every song playing was one my wife and I knew — all from our era. I started wondering: did they know? Could the music be picked for the people in the room?

That question became a side project. Baseball players have walk-up songs — a personal anthem that plays as they stride to the plate. What if everyone could have one? A camera watches you approach, estimates your age, picks a song from your era, and plays it on Spotify.

The Demo

Walk up to the camera and it'll guess your age, read your vibe, and play a song you'll recognize. Come back and it'll remember you and pick something different. The whole thing takes about five seconds.

Under the hood: OpenCV for face detection, ArcFace for recognition, Claude Vision for age estimation and song selection, and Spotify for playback. A laptop and a webcam — no special hardware.

The Talk

This isn't a talk about face detection. It's about innovation and creativity in the age of AI coding — what happens when you have an idea, start building, and let the tools take you somewhere unexpected.

Along the way: lessons on large vs. small models (the biggest model is rarely the best one), where AI-assisted development excels and where it falls short, and why “good enough” accuracy changes your architecture. The system told my wife she was a 69-year-old man. The first version took 36 seconds to pick a song. I'll share what went wrong and what I learned fixing it.

How It Works

Detect

YuNet DNN finds faces in the live webcam feed

Recognize

ArcFace embeddings identify returning visitors

Analyze

Claude Vision estimates age, reads style, picks a song — one API call

Play

Spotify starts playback in about five seconds

What You'll Take Away

  • The biggest model isn't always the best one. A general-purpose vision model outperformed a specialized age estimation model at a fraction of the cost.
  • Measure before you optimize. A single timing log revealed the parallelization opportunity that cut latency by 51%.
  • AI-assisted coding is powerful but it doesn't replace thinking. It writes the thread pool code perfectly — once you tell it what to parallelize.
  • Start with a fun idea. The best way to learn is to build something you actually want to show people.