I build browser-based gaze tools for dementia and ASD screening research on the laptops clinics already have.
Measure
EyeTrax uses MediaPipe landmarks for camera access, calibration, smoothing, and the live gaze overlay. It is open source and works on an ordinary laptop. Field testing exposed head-motion leakage that never showed up when I tested alone, so I built a newer method designed to reduce sensitivity to head pose. That approach is described in my recent preprint; the live demo runs the open-source baseline.
324 stars · MIT · Python and JS
Model
EyeCI uses gaze traces from a set of images to produce exploratory screening scores. Field sessions showed that head movement was leaking into the predictions, which had not appeared in my solo tests.
Deploy
NeuroSight carries EyeCI into field research. Project Argus supports dementia-focused screening research in clinics and community events. Project Iris adapts the same system for ASD-focused research with funding from Xunfen Biotech. I also built the collection platform itself: the console research staff use to run sessions, track quality, and export data.
2,095 session-level runs evaluated through Project Argus
Side projects
I made both because I wanted them in my own setup.
Portfolio index 01-02 Selected 2024-Now
Reddix
I built Reddix in Rust because I wanted to browse Reddit from the terminal. It has picked up 959 stars since, so I maintain the async code and caching as the project grows.
Open Reddix
Mistake.nvim
I write most things in Neovim and missed having autocorrect. Mistake.nvim started as my fix for that, and about 20,000 entries and 56 stars later, other people use it too.
Open Mistake.nvimAbout
I'm CK (Kevin) Zhang, and I'm heading to UCLA this fall to study cognitive science. For the past two years I've been trying to find out how much useful screening data can come from a laptop webcam instead of lab equipment.
Research-grade eye trackers are out of reach for many clinics, while laptops are already there. I keep the system browser-based and make the tasks simple enough that a participant can just watch the screen. The tradeoff is noisy data, and I still need a much larger data set before I can make stronger claims.
If you're working on gaze tracking or cognitive screening, send me a note. I'm happy to talk about the parts of the research that are still unresolved.
ck.zhang26@gmail.com