Stuff I’ve made

Club builds, class projects, research of my own, and a couple of things that are still only proposals. I’ve said which is which. Places I’ve actually worked are over on Work.

Hardware

  1. Sawdawg: 12 lb Walker Combat Robot

    2026 to nowMIT Combat Robotics, Exec Member

    Chassis design on a 12 pound combat robot.

    • CAD
    • Machining
    • Electronics

  2. MASLAB Autonomous Robot

    2026MIT Mobile Autonomous Systems Lab3rd place

    Five of us built a robot from nothing in under a month. It found its own way around the arena using computer vision, and stacked cylinders with a belt drive mechanism. I worked on mechanical design, manufacturing, and electronics. We came 3rd in the class competition.

    MASLAB 2026 Competition Highlights
    • Mechanical Design
    • Electronics
    • Systems Engineering

  3. Formula Car Lap Timer Mount

    Sep to Dec 2025MIT Motorsports, Chassis and Testing

    Worked on a lap timer mount to improve telemetry. Soldered and did design work.

    • Onshape
    • CAD
    • Manufacturing

  4. YUUMI: Drone Sensing for Steel Reuse

    2026early stage proposal

    A drone that carries RGB and temperature sensors alongside hyperspectral imaging, so structural steel can be surveyed and reused rather than scrapped. Routing algorithms decide where it flies. I’m leading the technical side and we’re working toward an MVP. Nothing is built yet.

    • Hyperspectral Imaging
    • Drones
    • Onshape

  5. Science Olympiad Build Devices

    2022 to 2025

    Devices for the build events. Deterministic robot built on Arduino and C++, built in a team environment.

Software

  1. Steel Lifecycle Mapping

    2023 to 2026NYU Urban Modeling Lab

    Structural steel gets scrapped and melted down when a lot of it could simply be reused. I mapped where American steel actually comes from and where it goes: mills, fabricators, recyclers and ports laid over the road and rail network, with each project sized by cost. On top of that I wrote a multiple origin, multiple destination router based on Dijkstra’s and ran it across 2,500 data points. Sending steel straight from one project to the next came out around ten miles, against hundreds of miles for the trip through a recycler. An abstract from this work was accepted to IABSE 2026.

    Box plot on a logarithmic mileage axis comparing two routing strategies, showing direct reuse producing far shorter trips than routing through a recycler.
    Routed trip lengths on a log scale. Going project to recycler to project runs into the hundreds of miles. Direct reuse between projects sits near ten.
    • ArcGIS
    • Python
    • GIS

  2. Phantom Traffic Jam Detection

    sole author research

    Traffic jams that form with no crash or bottleneck causing them are hard to see coming. I built Bayesian networks to catch them in the US DOT’s NGSIM trajectory data. After clearing out outliers and binning the features, the best models cleared 90 percent on every metric, with accuracy and recall both just under 94.

    A Bayesian network diagram with nodes for movement, lane, position, headway and vehicle dimensions, all linked to a central acceleration node.
    The learned network over NGSIM trajectory variables, all conditioning acceleration.
    A flow diagram showing the NGSIM dataset and a Bayesian graphical network feeding data preprocessing, then model training, then model evaluation against accuracy, recall, precision and F1 score.
    The pipeline: NGSIM data and a graphical network feed preprocessing, training, then evaluation.
    • Bayesian Networks
    • NGSIM
    • Machine Learning

  3. Construction Injury Prediction

    2023NJIT SCIIS Lab

    Construction sites hurt people in ways that look predictable once you have enough records. I built Bayesian network models over 4,847 OSHA incident reports that predict whether an incident turns out fatal. I vectorised the incident text, clustered it, stripped the outliers, then bagged a naive network and a tree augmented one into an ensemble. It hit 92.0 percent accuracy and 91.1 percent recall.

    The accuracy is not the interesting part. It ranks fifth of the seventeen studies I tabulated, and the four above it are all black boxes. On a real site a model that tells you which factors drove the call is worth more than one a fraction of a point better that cannot explain itself. I read and tabulated about thirty papers to work that out.

    Horizontal bar chart comparing reported fatality prediction accuracy across seventeen studies, with this work at 92.0 percent, ranking fifth and highest among the interpretable white box models.
    Accuracy against the sixteen prior studies in my literature table that reported it. The four above are black box models. The only other Bayesian network sits ten points lower.
    A method flow diagram covering dataset retrieval, text vectorisation, clustering and outlier removal, ensemble learning training guided by a knowledge graph, and evaluation metrics.
    The method: retrieve the OSHA data, vectorise and cluster the text, strip outliers, then train an ensemble guided by a knowledge graph.
    • Python
    • scikit-learn
    • Bayesian Networks

  4. LEO Satellite Commands and Data Handling

    2026 to nowMIT Satellite Team

    I’m writing the low level communication procedures for a satellite that will test a new kind of propellant mixing baffle in low Earth orbit.

    • C&DH
    • Comms

  5. AOVS: Variable Speed Limits

    2024research proposal

    A proposal to set freeway speed limits with machine learning you can actually interpret, using Bayesian networks and white box LSTMs trained on NGSIM data and checked in a traffic simulator.

  6. 2D vs 3D Reaction Time Study

    2022

    A class experiment with 16 people, comparing how quickly they reacted in a 2D aim trainer against a 3D one. A two sample t test found no real difference between them.