Portfolio (WIP)

• Cavalier Racing — Indy Autonomous Challenge

  Team Member
  Recently, I wrote code to process LiDAR point clouds of racetracks, to extract surface normals from the roadway surface. (This was neccessary to race at Laguna Seca (first road course), pictured.) Then I extract the norms along planned race-lines, and to load them into a k-d tree, which serves pitch and roll in real time on the car. This allows the gravity component to be accurately subtracted from accelerometer readings, providing better inputs to the localization EKF.
  
  I also recently wrote code to live-compare the expected/optimal with actual time for each turn, to help identify which areas to investigate after the run.
  
  Current Project: The problem is that the vehicle follows a pre-computed race-line, and we modify speed during testing by scaling or setting a maximum speed. However, we want to test by setting maximum (lateral and longitudinal) tire forces. My project is to compute optimal trajectories and velocity profiles constrained by tire-force limitations, on demand.

• Qwen Attention Visualizer

  I published a notebook to generate tensor files that contain the attention paid by every token to every other at every layer in small dense Qwen3 models, and an interactive visualizer to explore them. This is useful for understanding how these large language models work internally. I also published several example tensor files to play with in the visualizer.

  • Small (400KB, 4B Model)
  • Medium (5 MB, 0.6B Model)
  • Large (20MB, 4B Model)