Led multiple ML projects. Researched, developed, tested, trained, and deployed machine learning models and applications.

• Led research project on LLMs, IT Modernization, and code understanding
• Led research project on training object detectors for neuromorphic cameras
• Maintained internal GPU servers and services on OpenShift and Linux servers
• Developed and open sourced fire simulator to be used in RL training for wildfire mitigation
• Developed and open sourced LLM IT modernization library
• Developed agentic LLM backend prototype for government sponsor application
• Directly deployed trained models on government sponsor systems

Wrote, simulated, synthesized, and implemented FPGA code in VHDL using ModelSim, Synplify Pro, Vivado, and Libero Designer. Designed PCBs in KiCAD and tested hardware.

Investigates using LLMs to modernize legacy government code by focusing on code-chunking methods to overcome input limitations, showing that LLMs can effectively partition code and generate high-quality documentation.

Presents preliminary findings on the impact of documentation on LLM comprehension of legacy code, leveraging multiple-choice question answering (MCQA) for evaluation.

Investigates using LLMs to generate documentation for legacy code in MUMPS and ALC, proposing a prompting strategy and evaluation rubric while highlighting the limitations of current automated metrics.

Empirically analyzes how code documentation quality impacts the code generation and understanding capabilities of Large Language Models (LLMs). It reveals that incorrect documentation significantly hinders LLMs' code comprehension, while incomplete or missing documentation has no significant impact.

Presents a reinforcement learning approach to wildfire mitigation in simulated disaster environments. We release two software libraries, SimFire and SimHarness, to facilitate future research in this area.

Investigated the vulnerability of machine translation systems to adversarial attacks constructed with limited information. A novel attack method was proposed that generates perturbations using paraphrases and evaluates their impact on meaning preservation and translation degradation across various language pairs and systems.

Proposes a methodology to evaluate the vulnerability of unmanned aerial vehicles (UAVs) to adversarial machine learning attacks by analyzing potential attack vectors at each stage of UAV operation.