Face detection: 9ms Face recognition: 130ms per person Multi-face tracking that actually works Runs locally with zero data transmission (with an option to stream feed to PC, but it's entirely up to you)

The hardest part wasn't the AI - it was figuring out how to make anything work on hardware so new that most examples don't compile. With examples that don't cover the whole span of what developer need in the real world. Spent months reverse-engineering and debugging cryptic NPU errors, hacked and slashed my way untill things worked - ST might learn new things about their own hardware from how I abused some of it. Includes full source code, build pipeline, model conversion scripts, and documentation written by someone who actually used it. No enterprise sales pitches - just working code for embedded AI. Have fun and :)