Starting a notebook on edge perception

I’ve spent most of my career moving computer-vision systems from a research prototype to something that actually runs — in real time, on a device with a fixed power and memory budget. That gap, between a model that works in a notebook and one that ships, is where most of the interesting engineering lives. This is where I’ll write about it.

What to expect

Short, practical notes rather than polished essays. Topics I keep coming back to: real-time object detection, visual-inertial SLAM, multi-camera calibration, and the compression work — quantization and pruning — that makes any of it fit on edge hardware like the Jetson Orin.

Expect concrete trade-offs: where INT8 quantization-aware training helps and where it quietly costs you recall, how to build a model-error taxonomy that actually drives data curation, and the unglamorous details of getting a pipeline to hold its latency budget under load.

Why write it down

Mostly because I keep re-deriving the same lessons. Writing them down makes them reusable — for me, and hopefully for anyone else shipping perception to constrained devices.

More soon. In the meantime, the code lives on GitHub.