Camera-guided exercise apps split into two architectures. One streams your video to a server, runs the analysis there, and sends results back. The other ships the model to you and runs it where the video already is: on your phone or laptop, inside the browser. KineTrue takes the second approach, and the difference is worth understanding, because it changes both what you feel during a session and where your video ends up.
Real-time feedback has a physics problem
A camera running at 30 frames per second gives the software about 33 milliseconds per frame. A round trip to a distant server typically takes longer than that on its own, before any analysis starts, and home Wi-Fi under load makes it worse. A cue like "keep your knee aligned" is only useful mid-rep, so cloud processing isn't just slower for this job; it misses the moment. On-device models were built for exactly this constraint: in-browser benchmarks show the BlazePose family tracking a full 33-landmark pose at 30 or more frames per second on ordinary phones and laptops.
Privacy by architecture, not by policy
A privacy policy says what a company promises to do with your video. Architecture decides what it can do. Video that is processed on your device and never uploaded can't sit in a server bucket, can't be retained by mistake, and can't be exposed in a breach of someone else's infrastructure. Reviews of digital health systems make the same point in general terms: centralising sensitive data increases exposure risk, and keeping it local reduces it. For footage of you exercising in your living room, that's the difference that matters.
What actually leaves your device
Results, not footage. During a session the analysis produces small, structured outputs: repetition counts, movement-quality signals, and your pain reports. Those go to your account so your therapist can review them; the frames themselves don't need to travel. The size difference is enormous (streamed exercise video is measured in gigabytes per month, results in kilobytes), which is also why sessions hold up fine on modest internet connections.
Your browser got quietly good at this
None of this requires an app store install. Browser machine learning crossed a threshold over the last few years: WebAssembly gained vector instructions and multithreading in 2020, making in-browser inference several times faster, and as of late 2025 WebGPU, which gives web pages direct access to your device's graphics hardware, is supported in Chrome, Edge, Firefox, and Safari. The device you already own is enough.
On-device processing covers the analysis of your camera feed. Your session results are still saved to your account, deliberately: a monitoring product only works if your therapist can see how you're doing.