Around some 5,000 days ago, I built this.
This was me experimenting with what you'd now casually file under "AI", but back in 2012 there was no such label to hide behind. It was just a Kinect, a depth camera, and a lot of code. My first gesture-controlled game, "Dish Breaker": swing your hands, smash the flying plates.
The Kinect gave you a rough skeleton, about 20 joint positions, streamed at 30 frames a second. Raw, that data was messy: the joints jittered, the hands drifted, and the faster you moved, the noisier it got. The real work was never the game. It was smoothing that stream, deciding what actually counted as a "hit" when the coordinates were never exact, and keeping the whole read-process-respond loop fast enough that a swing connected the instant it landed.
Here's why I still think about it. Everything we now call AI, the models, the predictions, the "intelligent" systems runs on top of exactly this: reading the physical world through a sensor, and turning noisy, imperfect signals into something a machine can trust and act on. The intelligence is only as good as the sensing underneath it. A brilliant model fed bad sensor data is just confidently wrong.
That's the whole reason I build environmental sensors today. Back then it was a hand hitting a plate. Now it's air quality on a live dashboard. The output changed. The problem underneath, clean, reliable sensing in real time never has.
Get the sensing right, and the AI takes care of itself. Get it wrong, and no amount of AI saves you.
https://www.youtube.com/watch?v=UvyOSzGkjHk
