Track Record
I 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.
Four 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.
Across 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.
I 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
Granted and published US patents across eye tracking, hand tracking, spatial mapping, and data security.
Shared AR requires devices that are rarely equals - headset vs phone, different sensors and compute - to agree on where they are. This patent enables cross-localization into one 3D map via feature descriptors generated at multiple fidelity levels and converted between formats in transit, so each device localizes with the representation it can afford. It is the plumbing underneath colocated multiplayer and shared world-locked content.
Hand tracking needs depth sensing that is always ready, but an always-on time-of-flight sensor drains a wearable in hours. This scheme idles the sensor in a low-power cadence of sparse depth and amplitude frames - just enough to notice hands entering the interaction volume - and switches to a high-accuracy sequence only when an activation condition fires. Full-fidelity sensing exactly when needed, near-zero cost the rest of the time.
Conventional eye trackers image the pupil with a camera and several IR LEDs - power a wearable cannot spare, at video frame rates. This design tracks the eye by the laser speckle pattern its surface creates: one coherent source and one bare, lensless detector, with frame-to-frame optical flow of the interference pattern resolving micron-scale movement at rates up to ~15,000 fps. An eye tracker that fits an AR glasses power budget.
AR glasses have nowhere good to put an eye camera - direct line-of-sight to the eye means hardware in the wearer's view. This apparatus images the eye through a waveguide: a diffractive element in front of the eye couples the image in, total internal reflection carries it along the substrate, and a second element ejects it to a sensor hidden at the frame's edge. The same trick the display uses to deliver photons, run in reverse.
Earlier security-domain work: files are tagged with identifying elements at creation inside the internal network, so any file later observed crossing to an external network can be traced to its origin. Data-exfiltration detection built into the file lifecycle rather than bolted onto the perimeter.