In this video, I design a Squid Game–inspired “Red Light, Green Light” system the way you’d build it in the real world: as a real-time computer vision pipeline, not just a model demo. I start from problem formulation—what exactly counts as a “movement violation,” how strict the decision should be, and what latency/accuracy trade-offs you can actually afford when people are moving in a crowded scene. From there I break down the core building blocks: pose estimation with MediaPipe, multi-object tracking with DeepSORT, and a hybrid ML + rule-based layer that turns noisy vision signals into decisions you can explain and debug. The main takeaway is system design thinking: functional requirements (detect motion reliably, track identities, handle occlusions) and non-functional requirements (low latency, scalability to 100+ participants, multi-camera inputs). I also cover how I’d evaluate this end-to-end with offline metrics (tracking stability, pose confidence, false positives) and online metrics (latency, throughput, violation rate drift). Finally, I map out architecture patterns for scaling—streaming video ingestion, per-camera workers, synchronization strategies, and monitoring—so you can take the same blueprint and apply it to real-time CV products beyond this fun Squid Game framing.