google-ai-edge/mediapipe — reverse-engineered prompt

Build me a MediaPipe style project for on device machine learning that can handle live and streaming media without sending the input off the device. I want it to work across desktop, web, Android, and iPhone, with ready to use solutions for vision, audio, and text, plus a lower level framework for chaining processing steps together into custom pipelines. Please include a few example apps people can actually run, make the core parts reusable and customizable, and keep the focus on real time performance for things like camera, video, and audio streams. It should feel like an open source toolkit that can be used as is or extended for custom apps. If there are older example solutions in the repo, keep them available but make the newer solutions the main path. Please wire up build and setup so the examples are easy to try on desktop first, and look up the current MediaPipe docs online if you need to fill in any missing details.