Track Record
Visual Positioning System - Meta Quest
2020-2024I initiated Meta's Visual Positioning System, served as its chief architect, and managed the program from applicative research to production at more than 1M daily active users - Meta's first AI/CV cloud deployment for wearables. Along the way I directly led 20 researchers and engineers and tech-led a 90-person organization across the US, EU, and Israel. VPS builds a world model of the user's environments - the first consumer product ever to do so - anchoring shared world-locked content, recognizing spaces across lighting and furniture changes, and cutting boundary-recall failures in half across 100M+ space recalls. The system now extends beyond Quest to Ray-Ban Meta and Meta's upcoming wearables.
Perception & Sensing - Magic Leap
2016-2020Four years at Magic Leap, three roles of widening scope. I led the global research effort behind the company's wearable eye-tracking system and founded its primary synthetic-data team. I then tech-led a state-of-the-art machine vision sensor from research to mass production - a 30-person team plus a multitude of vendors, spanning a CMOS image sensor, a computer vision ASIC, imaging and diffractive optics, algorithmic pipelines, factory calibration, embedded firmware, and the perception SDK. I finished as Director of Engineering: final owner of the system architecture and specs of the company's XR wearable, leading a global group of 30+ engineers and architects and directing the work of 1,000+ engineers - a role I took over from a company co-founder, reporting directly to the CTO. The public teardowns document the hardware.
Ray-Ban Stories & Ray-Ban Meta
2021-2024Across both generations of Meta's smart glasses program I contributed on the camera and computer-vision side - from Ray-Ban Stories' first-person capture to the multimodal AI that lets Ray-Ban Meta reason about what the wearer sees.
Synthetic Data
2016-2024I led synthetic-data programs at both Magic Leap and Meta - simulation-first training for perception models years before it became standard practice. At CVPR 2022 I was an invited speaker and round-table panelist at the BMTT workshop 'How Far Can Synthetic Data Take us?', alongside Jitendra Malik, Kate Saenko, and Deva Ramanan.
Patents
10 granted US patents in 5 families across eye tracking, hand tracking, spatial mapping, and data security.
Power-efficient hand tracking with time-of-flight sensor
Hand tracking needs a depth sensor that is ready when a hand appears, without running it at full accuracy all the time. This method runs a time-of-flight sensor in a low-power sequence of a depth frame plus at least one amplitude frame, and switches to a high-accuracy sequence of repeated depth frames once an activation condition is met.
Eye-imaging apparatus using diffractive optical elements
An eye-imaging device for AR and VR head-mounted displays that uses a substrate as the light path. Two offset coupling elements sit on the substrate: the first deflects light so it travels inside the substrate by total internal reflection, and the second deflects it back out.
Methods and systems to allow three-dimensional map sharing between heterogeneous computing systems, cross-localization, and sharing content in three-dimensional space
Artificial-reality devices that track the world differently still need to share a map of it. Here one device derives lower-detail feature descriptors from a high-detail set to track its own surroundings, and transmits the high-detail set so it can be converted into the different descriptors a second device uses to track the same environment.
Method and system for eye tracking using speckle patterns
Following speckle from frame to frame measures how the eye moves, but gaze also needs an absolute starting point. This method finds that starting position from a pre-calibrated speckle map of the user's eye, then tracks movement from it by following the speckle patterns that light reflected by the eye forms on a detector.
Computing system and method for identifying files transmitted to an external network
Earlier security-domain work on telling when internal files reach an external network. An insertion device tags selected files with a data file element, and when a tagged file is opened, read or processed on an external computer, that computer reports to a monitoring server.