I enjoy making things. Here are a selection of projects that I have worked on over the years.
Building an end-to-end inference stack for open-source LLMs and documenting lessons learned.
See the presentation below for strategies to optimize LU decomposition with MPI by blocking the data, building detailed data-dependency graphs, and executing in parallel wherever possible.
Dance Cube This project includes a 3D 8x8x8 LED cube connected and controlled via WiFi by an Android app. In addition to animated lighting, the phone camera runs an on-device ML model for pose estimation and projects the real-time movement of the person into the cube in 3D.
This project implements parallel and accelerated FFT variants in OpenMPI and CUDA, grounded in the mathematical and algorithmic foundations of the FFT and its historical development. It highlights the divide-and-conquer structure that reduces DFT complexity from O(N^2) to O(N log N), then develops the additional theory needed for parallel execution and addresses key optimization issues. Results are presented along with a discussion of known issues and potential improvements. See the presentation below for the technical walkthrough, and download the report for the full methodology and analysis. Project presentation (PDF).
Using a dataset that mapped EEG signal recordings to specific physical hand motions, we built a CNN that classifies the physical action by analyzing the EEG recordings. Project presentation (PDF).
