Look away and it notices. A webcam tracker that turns eye contact into an interface. If you lose focus, it responds in real time, proving that machine learning should help you stay present, not just log your distractions after you've already drifted.
Most of my machine learning work ends at a metrics table, and most of my frontend work has nothing behind it. This is the repository where those two meet, which makes it the closest thing I have to a statement of what I want to build.
The model runs on a live video stream, extracts face and gaze landmarks through MediaPipe, and derives an attention signal from them. The part I care about is what happens next: the signal is rendered back to the viewer in real time, rather than accumulated into a file that somebody reads afterwards.
That is a design constraint disguised as an implementation detail. A model whose output arrives a day later can be as slow and as opaque as it likes. A model whose output is on screen while you are still sitting in front of it has to be fast enough to feel immediate, and legible enough that a person can tell when it is wrong.
The code is finished and running. The full case study is still being written, specifically the accuracy characterisation: how the attention signal behaves across lighting conditions, head poses and distances, measured rather than asserted.
I would rather leave that section visibly incomplete than fill it with numbers I have not earned yet.
Think of the motion-capture suits actors wear in movies (with the little white balls all over them). This does the exact same thing, but completely invisibly using a normal webcam. It maps exactly where your head and eyes are looking.
This means the AI is "thinking" fast enough to keep up with real life. Instead of recording a video, sending it to the cloud, and waiting 5 minutes for a result, the AI processes your eye movements instantly on your own laptop so the screen reacts the second you look away.