Useful AI begins with an unreliable signal.
A face image, a tiny reflection in an eye, watch measurements, or a photo of a blood sample can all contain useful information. The engineering challenge is making that information dependable enough to act on.
My work across computer vision and healthcare follows the same path: isolate a useful signal, understand its uncertainty, design for the device it must run on, and turn the result into a clear user action.
I am an inventor on 15 patents. This article explains six families with public records and links; filed work without a public document is not described as published.
Make gaze tracking work outside the lab.
High-end eye trackers can use infrared cameras, carefully placed lights, and controlled conditions. Consumer devices do not always have that setup. These inventions attack different parts of that gap.
Estimate gaze on ordinary devices
Fuse information from the whole face with the eye that is most visible. The model keeps head pose and fine eye detail together while remaining suitable for mobile hardware.
Track with one corneal reflection
Traditional geometry may depend on several reflections from infrared lights. Using one reflection reduces hardware complexity and keeps tracking possible when movement hides the others.
Calibrate once, simplify normal use
Use controlled calibration to learn person-specific eye geometry. During normal tracking, estimate gaze without requiring the same known light-source setup.
Let eyes handle distance and hands handle precision
Gaze quickly identifies a region across a large screen. A mouse, touchpad, or pen finishes the precise selection, avoiding the jitter of eye-only pointing.
Turn passive signals into timely guidance.
Healthcare data is useful only when the full path works: collection, quality control, modeling, interpretation, and a clear next step for the patient or clinician.
Continuous signals from wearables
The digital-biomarker patent describes using signals from devices such as watches and phones to model correlations with blood characteristics relevant to cardiorenal metabolic conditions. The goal is not to present a watch as a laboratory; it is to make changes between clinical measurements more visible.
Read WO2025106664A1 ↗Check an at-home sample before it travels
The sample-card patent uses image processing to judge whether an at-home biological sample is likely usable. Giving feedback before shipment can avoid waiting days for a laboratory to reject a poor sample.
Read WO2023091455A1 ↗The last mile changes the research question.
Across vision and healthcare, the work moved from algorithms to complete loops: sense, estimate, validate, and respond.
Patent records and research foundations.
These links lead to the public documents behind the explanations above.