Andrew Boessen

A 3D Gaussian Splatting library built in C++ and CUDA, providing a high-performance, lightweight implementation for both training and real-time rendering of photorealistic scenes.

Neural network approach for modeling interactive game environments using Vector Quantized Variational Auto-encoder (VQ-VAE) and Spatio-Temporal Transformers. Trained on Atari Skiing gameplay data. A detailed technical report is available in the linked repo.

PerfectRep is a 3D pose estimation model tailored specifically for powerlifting analysis. It allows for precise tracking and analysis of lifter's movements to ensure perfect form and technique.

A tiny entirely unoptimized autograd engine with gpu acceleration built with CUDA and C. A small neural network library is included to trained multi-layer perceptron models with gradient descent by backpropagation.

Efficient CUDA implementation of "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention". This uses warp primitives and Tensor Cores for optimal performance.

EagleEval is a platform designed to empower Boston College students with valuable insights into teacher and class ratings. Used by 5,000+ students. Explore at eagleeval.com.

In place sorting of arrays using bitonic merge sort. Built using CUDA to optimize performance for GPU execution. This takes advantage of CUDA's warp shuffle operation to sort numbers in place and in register memory.

Re-Clip is a mobile app that leverages LangChang, MoviePy, Stable Diffusion, and Swift to summarize and generate short-form video content about research papers and GitHub repos through the use of an automated AI pipeline. Built at HackHarvard 2023.