Real-time room awareness, running entirely on a 15 MB footprint

Picture a smart camera that instantly knows how many hands are raised, who's standing, who's sitting, and the moment someone rises from their seat, all processed in real time, on the device itself, with no cloud, no personal identification, and no compromise on privacy.

That's the challenge we took on: building an AI model lightweight enough to run entirely on CPU-only embedded hardware, within a footprint smaller than a phone photo, while still delivering real-time performance in busy meeting rooms and classrooms.

Rather than relying on heavy, generic computer vision, we designed a purpose-built pipeline: compact pose estimation focused on the upper body, paired with simple geometric reasoning to infer gestures — a raised hand, a change from sitting to standing — without ever tracking or identifying individuals. The result runs comfortably within the device's tight memory and processing budget, delivering room-level insight through a clean, ready-to-integrate API.

Proof that powerful AI doesn't need powerful hardware, just the right approach.