Right now I'm using a vision library to detect head height which was good enough. I went down a tangent where I hooked it up to my Claude Code instance to take a screen shot and have Claude Code assess how bad my slouch was. Claude would watch a folder for screen shots, read it in, and if it detected bad posture, write to a file the program was watching to adjust blur.
I did this weird work-around so I could use my Claude Code subscription as opposed to the API.
Anyways, it was too slow and Claude was a bad judge of slouchiness. Head height works well enough!
I'll clean this up.