Abstract
<sec> <title>UNSTRUCTURED</title> <p>Most consumer movement-analysis products, including our own shipped system, still send a full video file to a server before anything useful happens with it. That design choice made sense five years ago, when running a real-time pose model outside a data center wasn't really an option. It doesn't make as much sense today. Google's BlazePose model has run at real-time frame rates on ordinary phone hardware since 2020, and browser-side inference through WebAssembly has closed most of the remaining gap between "runs on a phone" and "runs well enough to matter." This paper lays out a proposed architecture in which pose extraction happens entirely on the user's device, and only a small, numeric sequence of body landmarks -- never the video itself -- ever reaches our servers. We describe the current shipped pipeline honestly, including its actual bandwidth footprint measured from production, walk through what changes under the proposed design, and estimate the resulting reduction in data transmitted using a transparent, disclosed methodology. We are careful throughout to separate what we have measured from what we are projecting, and we close with a concrete plan for validating the proposed pipeline against the one already running in production before either one is presented as final.</p> </sec>