Projects Archive

At Cornell’s Collective Embodied Intelligence Lab, under Kirstin Petersen, I explored stigmergy: coordination through persistent changes to a shared environment. Bees leave pheromone trails in air. I wanted to know what that idea could look like in water, where connectivity is unreliable and the environment can carry a signal.

The bubble pattern was the message. Its timing, speed, and formation could represent a bit, a location, a direction of travel, or another shared state. I wrote the distributed processing and vision algorithms that made that loop usable, including image segmentation and tracking work that reduced processing time from roughly one second to 150 milliseconds. With that feedback loop in place, the agents showed meaningful multi-agent behavior without ever receiving a direct message from each other.

While the outcome of this particular investigation was incorporated into a larger stream of research yet to be published, the project was a first exposure to swarm robotics and the potential for low-cost high-volume robots to be placed in harsh or restrictive conditions.

I built Visions while returning to Kotor in Montenegro between the end of high school and my second year of college. COVID had emptied the historic old-town, with cruise ships and buses from up and down the Balkans missing. Tourism, accounting for 30%+ of Montenegro's GDP, needed its attractions to stand out more in order to attract a limited crowd. I was already deep into AR/VR at the time, and experiencing history had always been the part of travel I cared about most.

For VantageVR, I worked with an asset developer to turn the researched story and scene design into a Unity experience packaged as an APK for Oculus Quest. The experience was launched for paying customers in Perast and aboard boats travelling to Our Lady of the Rocks. I validated the reconstruction through sustained work with local officials and museum curators, plus a lot of primary and secondary research on Perast.

KotorAR extended the idea to a phone. Visitors could hold up a phone at an attraction, see an AR reconstruction of its earlier form, then follow guides and context into the story around it. I owned the product and technology architecture around the app, including local-business and mass-transit integrations. The asset developer built the AR scenes; I made sure the experience connected those scenes to the city around them.

The project was backed by Montenegro's national tourism organization and was discussed in the country's Senate during the COVID era as part of the conversation around tourism technology infrastructure. It came as small businesses suffered a lack of visibility in an increasingly restricted old-towns, and is a concept that has since seen adaptations in many major cities worldwide.

Tik Tok is an active bio-hybrid humanoid project spun out of Professor Andy Ruina's work at Cornell. I helped design and fabricate the PCBs that captured and relayed rotations at the knees, hips, and ankles. The boards homed multiple subsystems in a fixed packaging, making layout and assembly challenging. Here I learned to use stencil and validation tooling for SMD components, alongside signal integrity and mixed signal logic.

The payoff was seeing a multi-joint rollout start to resemble a gait. Nothing glamorous about a clean encoder trace, but it is a good day when the whole stack finally agrees on where the body is.

rotary encoder
  → Cortex-M7 embedded procesing
  → CANOpen / EtherCAT communication
  → Nvidia Jetson timestamp alignment
  → multi-joint gait rollout

When I joined Cornell Electric Vehicles, it was a team of ~20 and solely purposed for Shell Eco Marathon Americas energy efficiency competition. However, it had already been 10 years since the first mass-market Nissan Leaf (quite full-circle, this is the robotaxi platform of choice for Wayve!), and a few of us felt the team needed a new mandate. My first year was across our sapling Autonomy team, where traditional computer vision and SLAM led us to promising results in sim and virtual competitions (still COVID-era). In parallel however, I learned more about PCB design, and leaned into building the first 'Electrical Automation' board. This board served as what industry calls a 'vehicle interface controller', switching and conducting integrity checks on actuation outputs/feedback from our onboard autonomy compute. This effort paved the way for our eventual L1 autonomy demonstration: Autonomous Emergency Braking. The team has since championed the autonomy stack well past L1, and on vehicles of multiple scales, making autonomy a first-class pursuit of its members.

Most of my engineering time went into 7 revisions of our 2kW BLDC Motor Controller (or 'Electric Speed Controller'). I experimented with everything mixed signal and power here: different power layouts, commutation patterns, heat dissipation, and (most frustratingly) load protections. One of these Motor Controllers ran our 150kg car and driver around the paddocks of the iconic Indianapolis Motor Speedway!

Outside of electrical and autonomy development, the workshop became a second home for me in college, and my teammates a second family. I learned a ton, from carbon fiber resin layups and machining extrusions, to aero profiling and low-latency telemetry. From a birds-eye view of technical roadmaps running all-hands, to the minutia of subteam struggles, some of my best (and most challenging) memories are with the team.

Many high schools have teams where students explore engineering, but at the International School of The Hague the team was more serious. The Netherlands has a sterling record of high school and university teams competing at the frontiers of energy technology in mobility, particularly across battery-electric and hydrogen fuel sources. The Silverbacks Engineering team I led still holds one of the strongest efficiency results recorded by a high-school team in the Shell Eco Marathon Europe competiton.

MARTHA ("Monitoring and Restriction Tracking Hotspot Application", but more importantly named for Cornell President Martha Pollock) was a hackathon solution to the problem of avoiding hotspots for unsafe activity around campus. Dining halls, frat row, and campus buildings were the most obvious places to avoid, but how could you know if your walking route to classes was packed?

While lockdown ended before a solution like this became necessary, Fall 2020 was the perfect time to build and demo the capability of such a service. There are potential post-COVID applications of such opt-in tracking tech, like crowd management for venues, city organizers, or even for public services to track usage.

SAFE was a concept inspired by the heatstroke of an elderly relative in their home in India. It sounds like an isolated incident, and not one we usually hear of in the West, but is surprisingly common in the East, particularly among the elderly. The research is that bodily thermal regulation capability reduces with age, as well as the ability to accurately gauge enviornmental conditions. Combined with rising temperatures globally, and the liklihood of households in developing countries to lack central climate control, SAFE was intended to act as a canary for unsafe enviornmental conditions.

To trial the system, I took a handful of pods to India and presented to an esteemed geriatric physician, who oversaw my successful trials in a handful of old-age homes. Should the system have been taken forward and productionised, it could have application in senior homes and hospitals everywhere.

One interesting failure mode I noted from the India field trials was due to missing inputs (mainly sensor drop-outs), causing the model to fail forward passes and provide inaccurate risk assessments. This prompted my research on neural network input nullification for healthcare applications, and proposal on a heuristic to reduce the impact of missing data. The resulting paper was presented in a talk to University of Amsterdam faculty.