A decade shipping computer vision into production, much of it converted and running on-camera and on-NPU for enterprise clients in Japan. Mobile, cloud and agent work across Asia, Europe and Africa alongside it.
One number without the other is a claim nobody can check, so each system reports the flattering framing and the unflattering one side by side and says which is fair to which measurement. The evidence below is what that discipline produced; it grows, the standard does not.
Over the past decade I've shipped software for enterprise clients across Asia, Europe, and Africa: real-time vision systems on factory floors, encrypted mobile deployments, cloud pipelines, interactive 3D applications, and autonomous automation that runs businesses day to day. Seven of those years were edge AI for Japanese enterprise clients — models converted and shipped onto vendor NPUs and on-camera accelerators, where the toolchain turns out to be most of the work. Whatever the field, my specialty is the same: building things that actually work outside the demo.
Most of my output belongs to the companies that paid for it — client deployments, proprietary architectures, internal tooling. I can say what the work was and which silicon it ran on, and nothing more specific than that. The repositories above exist because a claim nobody can check is worth very little, so the method is demonstrated on work I own outright.