Galton’s Goalie – robburke.net

The Spark

My son and I have been building Mark Rober’s CrunchLabs Build Boxes and they’re an absolute blast. For the uninitiated: Each month, Mark Rober and team introduce an engineering concept with a ~20 minute video.. and then you build something cool that demonstrates it hands-on. We just started Season 2, and the first box was all about the bell curve distribution – demonstrated with “Galton’s Goalie,” basically a desktop Plinko machine.

Watching the balls bounce randomly through the pegs and settle into a pretty decent bell curve distribution is mesmerizing. But somehow it wasn’t proving the point well enough.. and I couldn’t help thinking: what if we could actually PROVE it’s creating a bell curve? With computer vision layered on top, we could track every ball and visualize the distribution in real-time.

So I thought: what a fun vibe coding project! Let’s see if we can do that. Spoiler alert: We did.

Real-time ball tracking and histogram visualization

The Journey

This was my “OMG I just did 3 months of work in a couple of hours” moment.

Back in the day, I built Still Life in our Symphony engine – a signal processing engine with a computer vision component. We spent literally months building the signal processing and computer vision foundations. Detecting motion, tracking objects, creating particle effect trails.

Fast forward to 2026: I described the problem to Claude, we talked back and forth about the approach, and it baked the first working version in a matter of hours. Real-time ball detection. Histogram updates. Motion trails. The whole thing.

But the journey didn’t end there! I showed it to a colleague who asked if we could improve the UI. One conversation with Claude later, we had a complete PyQt5 interface with multiple visualization modes, calibration controls, the works.

The updated PyQt5 interface with visualization controls

Why Not Both?!

With Still Life, adding a visualization mode (long exposure, particle effects, etc.) was a significant decision. We’re talking days of careful coding, testing, integration. You had to commit, then grind.

Now, when Claude would suggest a few different visualization options, I could say “why not both?!” and both would be cooked that afternoon. Long-exposure mode. Motion trails with color cycling. Per-bucket sensitivity tuning. Why not!

What used to take days now takes minutes. The limiting factor isn’t implementation, it’s the idea itself.

But I’ve Done This Before

I wonder how important it was that I’d done it before. I knew what questions to ask Claude, and intuition for what CV approaches would work when it gave options. The vocab was already in my head.

Regardless – you don’t need months of CV experience to build something like this anymore. Just curiosity, willingness to iterate, and the ability to recognize when something’s working (or not).

This was the project that inspired me to start this whole “Side Quests” series. If I can knock out a real-time computer vision app in an afternoon, what else becomes possible?

Try It Yourself

The full code is on GitHub: GaltonGoalieViz

Watch an earlier version in action on YouTube, or check out the sample videos showing the system tracking balls in real-time:

Sample visualization in action

You’ll need:

Python 3.7+
A webcam
Something that drops balls through obstacles (Galton board, Plinko, whatever .. and if you’ve got nothing, just set up calibration zones, wave at the camera and make yourself a histogram)
About 5 minutes to get it running

How’s that for “Engineering Hustle”?! Hey Mark and Team CrunchLabs, if you ever stumble across this, thanks for your spirit and all the inspiration. ;)