Duration: Really scattered, mostly built over a single week in July 2026, testing ongoing
Motivation: I saw this video on Youtube about pianists wearing eye trackers and wanted to try it myself. Most notably I was interested in what I looked at when my hands were moving in two different directions, a use case not explored by the Youtube video.
Outcomes: It works pretty well when the thing the user is looking at matches the calibration state. For example, for the piano use case, I had to calibrate it with my laptop screen on my lap to match how my gaze is pointed downwards. It works really well if I look at anything on my laptop because, well, the calibration is done while looking at point on a laptop screen...
But yeah, didn't really spend too much time on the calibration algorithm, so that could definintely be improved. Also the world-facing camera FOV is quite limited compared to how much my eyes can actually see when my head is pointed any given diection
Instructions for building and using the device on my Github!
How I look when I'm typing cleanly (trial 1) vs making lots of mistakes (trial 2)
Context, I am out of practice in all the below videos, brace your ears...
Testing key-level accuracy by playing a simple scale
Playing non-memorized music. (Calibration may be a little off here due to me switching between two vastly different gaze directions)
This is a nice passage because the two hands are doing symmetric runs in opposite directions. You can see I have 0% faith in my left hand, 100% faith in my right hand lol
Another example of two hands going in opposite directions, the left-hand bias is less strong here
A passage with less eye movement than you might expect: the first page of La Campanella
A passage with a lot of right-hand movement, but no right-hand jumps
vibe coding
like oh my god holy vibe coding. Gemini >> ChatGPT
Pi camera setup + streaming
How to switch networks on a headless Pi setup