Projects

Graph Zero is a novel AlphaZero-style algorithm incorporating Graph Attention Networks for improved game state representation, implemented in pure CUDA and C++. Scaled training to 8 GPUs, optimizing HPC performance using NCCL, resulting in an algorithm that outperforms AlphaZero in TicTacToe and Connect4, with promising progress in Chess. A paper detailing this innovative approach is currently in progress.

A neural wake word detection system utilizing EEG and eye-tracking data from the ZuCo2 dataset, featuring a custom Streaming Transformer built in PyTorch for real-time accuracy chunk detection. Performance was further optimized through the implementation of OpenAI Triton Kernels for GPU acceleration.

VLM-o is a 3B parameter multi-modal vision model built for image analysis and object detection built from scratch using PyTorch. Based on the PaliGemma architecture with an added dense connector and fine-tuned with LoRA on the VQAv2 dataset.

BitDB is a lightweight skeleton database similar to SQLite, written from scratch in C equipped with various memory ops and db tools. BitDB utilizes Ruby for testing various database scenarios and edge cases. It features a natural language to code assistant named Ada, implemented using MLX for native mac support with a finetuned Mistral-7B model.

This project involves developing a machine learning application to convert text into Braille, a tactile writing system used by people who are visually impaired. The core of the project is a long short-term memory neural network model implemented using JAX and Haiku. The model is trained to predict Braille representations of text characters.

This project involves constructing a 1.75 billion parameter GPT-2 based model from scratch. Training was conducted using the NCSA’s HAL Cluster. Key features of the project include innovative data generation techniques and rlhf using ppo. A simple chatbot UI was built using Gradio, and a “phone-call mode” was implemented for a more personal experience, leveraging AssemblyAI and Google Cloud technologies.

Built a predictive model using tensorflow to accurately forecast the UEFA Champions League top goal scorer. substantiated by successful prediction of Kylian Mbappé. Executed data scraping and cleansing techniques due to the unavailability of public, free, soccer datasets. Constructed a specialized C++ data store using unordered maps, empowering a comprehensive real-time dashboard for visualization of model results using a Kubernetes ci/cd pipeline.