I am an AI Engineering contractor for Fortune 500 enterprises. My work focuses on modernizing large-scale ML operations. I've built robust forecasting and proactive customer intelligence models, while architecting scalable MLOps infrastructure for distributed training, high-volume inference, and elastic compute orchestration using cloud computing and Ray technologies.
I did ML Engineering work at Covar where I built and shipped a mission-grade defense AI agent — orchestrating document retrieval, code gen, and ML pipeline gen for real-time decision-making. In this time I worked with finetuning LLMs, and shipping RAG source attribution methodologies.
I was a SWE intern at Vanguard where I built full stack applications, wrote an NLP sentiment model built on scraped Twitter data for buy/sell recommendations, and built an AWS data ingestion system moving 3M+ data points per day.
PipelinePilot: a multi-agent LLM system for prospecting and personalized cold-email generation, with persistent per-prospect memory for stateful agents and retrieval-driven context assembly grounding outputs in real-time data.
LLM-Inference-Speculative: An efficient, open-source speculative decoding engine for large language models, implemented in Python with PyTorch. Supports fast inference for models like Qwen2.5 and Qwen3, multi-architecture support (Apple Silicon, CUDA, CPU), and modular pipelines for customizable speculative decoding.
GAN4Seg: a CLI tool for batch brain-MRI tumor segmentation, continued at UNC Hospitals. Nibabel preprocessing on BRaTS 2021, PyTorch GAN for tumor ID + segmentation, DICE 0.86. MONAI + Slurm for the training stack.
BallotProof: full-stack client-side ballot image verification — OpenCV.js feature extraction, PIL-generated sample ballots for testing, and an AWS DynamoDB + Lambda annotation pipeline. HackGT7 Top 8 + HackGT finalist.
AzureWare: A full suite diagnosis and monitoring system for Parkinson's disease. Android app serves as interaction mechanism with custom glove hardware built for tremor measurements. Test results are stored via AWS API Gateway and DynamoDB. Tremor intensity and frequency are measured via custom algorithm based on accelerometer data and vocal characteristics are analyzed via RNNs. Won $10,000 at the Intel International Science and Engineering Fair over the course of work.