Projects and research

My research measures how GPUs behave under limits: how power caps change the energy cost of training, and how close LLM decode gets to the memory-bandwidth ceiling.

  • Combined

    In Progress: Senior Capstone
    TypeScriptElectronReactNode.jsHonoSQLitetmuxgit merge-treeMCP

    My senior capstone, which I’m building with two co-founders. Combined is a desktop app for running Claude Code, Codex and pi agents in parallel. Each new session gets its own git worktree, branch and tmux-backed terminal, so agents can’t overwrite each other and sessions keep running after you quit the app. Combined trial-merges every agent’s branch against main and the other branches, so conflicts show up before anyone opens a PR. We’re now extending this to teams, so your agents’ branches are also checked against your teammates’. In a test with staged teammates on real pushed branches, it caught the exact file where their work would conflict.

  • decode-sol-gb10

    In Progress: Personal Project
    PythonvLLMPyTorchCUDANVMLNsight SystemsRoofline modeling

    During LLM decode, every token requires reading all of the model’s weights from memory, so memory bandwidth sets a hard ceiling on speed. I’m deriving that ceiling from first principles for an NVIDIA GB10 and measuring how close vLLM gets to it across model sizes, batch sizes, context lengths and precisions. I’m using profiler traces to explain where the remaining time goes and tracking the GPU energy cost of every token. In progress, with no results yet.

  • GPU Power Capping and Energy Efficiency

    Computer Systems Research, VT Institute for Advanced Computing
    PyTorchGPU power managementPython

    How much speed do you actually give up when you cap a GPU’s power, and is it worth the energy saved? I studied this by training 5 ResNet architectures (ResNet18 through ResNet152) on CIFAR-100 at 5 power cap levels from 150W to 300W, each with a fixed 300-second training budget. For each run I measured total energy, average power, joules per image, and energy-delay product. I also looked at how model depth affects throughput, energy use, and generalization: the shallower models overfit, and the deeper networks validated better. I presented this work at the Fall 2025 VTURCS Research Symposium, advised by Dr. Kirk Cameron.

  • Broadband Address Matching for BEAD Funding

    SPIN Lab, VT Department of Computer Science
    PythonGeospatial analysisFuzzy matchingParallel processing

    Federal broadband funding depends on knowing which addresses actually have service, but ISP records and the FCC’s national address map don’t line up cleanly. I built a Python matching algorithm that reconciles the two. It uses fuzzy string matching on addresses and then checks that matched locations are within 50 meters of each other. With data cleaning, hashing, and parallel processing, it matches thousands of records in about 30 seconds. It reached a 93.4% overall match rate, with about 58.7% of matches at high confidence, which supports more accurate coverage estimates for underserved communities. I presented this work at the Spring 2025 VTURCS Research Symposium, advised by Dr. Shaddi Hasan.

  • MuseMatch

    J.P. Morgan Code for Good
    ReactTypeScriptTailwind CSSExpress.jsFirebase

    I built MuseMatch, an alumni network for artists who have come through the Lower Manhattan Cultural Council’s programs, during a 24-hour hackathon. Artists can browse each other’s profiles and portfolios with a swipe-style interface, connect with one another, follow a community feed of alumni news and grant deadlines, and create and RSVP to events. I used React, TypeScript, and Tailwind for the frontend, Express for the backend, and Firebase for authentication and data storage.