Joining YouTube’s AI team this summer to work on efficient neural network architectures for video processing on custom accelerators. The pull is the constraint geometry: real-world video, real-world latency budgets, real-world hardware.
Six months at the Autonomous Intelligence and Robotics Lab, working at the intersection of physics-informed ML and computer vision for defense applications. The thesis was that learned models could improve VLF radiowave military communications by predicting ionosphere state more reliably than physics-only methods, and the same lab also needed deployable vision for UAVs and satellite constellations.
Ongoing undergraduate research with the Autonomy, Robotics, & Intelligent Algorithms Lab at Mines. Two threads, both heading toward a first paper: localization that fuses semantic segmentation with object registration, and novel embedder architectures for convolutional image neural networks.
Two-month focused build on a synthetic-data pipeline for industrial robotic arms. Ricoh’s robotics team needed labeled training data at scale for object detection on a Universal Robots arm, and physical capture wouldn’t get there. The work was end-to-end: photorealistic warehouse simulation, diffusion-based scene generation, and a YOLOv11 detector trained on the synthetic output and deployable on the actual hardware.
Embedded systems work at SEAKR (an RTX subsidiary that builds satellite electronics), spanning low-level boot software, kernel-level security mitigations, and an ML demo aimed at a DARPA RFP. The unifying theme was reliability under hard constraints, radiation, memory, certification, that you don’t run into in cloud-native ML.
Returned to Lockheed Space the next summer for a research-track role on satellite control. The brief was to use ML to make magnetic torque rod simulations “digital-twin” faithful and faster than real time, so the simulation can eventually be flown on the satellite itself.
First industry internship, the summer before senior year of high school. Built an augmented-reality app from scratch for viewing full-scale 3D models of Lockheed Space products, then extended it to a peer-to-peer network so multiple program executives could walk through the same model together.