Projects

Projects I've worked on and currently working on.

Professional Projects

AI and machine learning systems I've built in professional environments, focusing on defense and consumer applications.

Asynchronous End-to-End Post-Training Pipeline

Active

Jun 1, 2026

Architected an asynchronous post-training pipeline combining SFT and GRPO reinforcement learning on Modal serverless 8×H200 compute, training open-source models to closed-source-level performance for binary and firmware vulnerability research.

Python Modal GRPO SFT AWS ECR gdb AFL++ ASan

Designed and built a fully asynchronous pipeline that decouples trajectory generation from policy training via independent workers communicating through a shared volume, so each stage scales independently on Modal serverless 8×H200 nodes. Paired with a complete data and evaluation environment: data pipelines for firmware vulnerability datasets, reverse-engineering tasks, and ASan-verified binary targets; a custom agentic harness with real exploitation sandboxes (gdb, AFL++, ASan) deployed via AWS ECR; and a comprehensive benchmarking suite spanning agentic, long-context, and security tasks. Developed under a SOCOM CRADA.

4-Node NVIDIA DGX Spark Cluster for Distributed Training and Inference

Active

May 1, 2026

Architected and validated a 4-node NVIDIA DGX Spark cluster (Blackwell GB10) interconnected via Mellanox switching, enabling distributed training and inference for models up to 120B parameters.

NVIDIA DGX Spark Blackwell GB10 Mellanox RDMA PyTorch vLLM

Led the architecture, build-out, and validation of a 4-node DGX Spark cluster interconnected via Mellanox switching. Benchmarked RDMA throughput and long-context performance across the fabric, standing up distributed training and inference pipelines for models up to 120B parameters. Also drove hardware selection and technical strategy for a 32-GPU lab expansion, and supported three field exercises and demos showcasing edge inference capabilities.

Custom Reinforcement Learning Environment for Agent Training

Completed

Sep 15, 2024

Co-developed a custom OpenAI Gym environment in C++ and Python, integrated with a Department of Defense simulation engine to enhance multi-agent training for AI agents.

Python Ray C++ Docker OpenAI Gym PySpark

Developed a custom OpenAI Gym environment to train AI agents on complex, real-world scenarios. Integrated this environment with a Department of Defense simulation engine using data pipelines, leveraging Ray for distributed training, PySpark for data manipulation, and proximal policy optimization to optimize agent performance.

StudyDuck: Self-Learning Study Tool Leveraging LLMs

Completed

Dec 1, 2022

Co-founded StudyDuck, a web application that utilizes OpenAI Davinci models and Hugging Face Gensim to create a self-learning study tool.

React AWS Lambda Python FastAPI Huggingface OpenAI API

Spearheaded the design, development, and implementation of StudyDuck, a natural language processing web application designed to enhance ebook learning for students. The application leverages a serverless AWS Lambda architecture with Python, FastAPI, and Large Language Models (LLMs) from Hugging Face (Gensim) and OpenAI (GPT-3). StudyDuck enables users to upload PDFs, and the LLM generates learning content including questions/answers, summaries, highlighted content, and keyword definitions.