Research

Reasoning and Learning Group
Apr. 2025 – Present Atlanta, GA

• Architected a post-training pipeline to adapt open-source Large Language Models (LLMs) to Material Science domain tasks. Pipeline leverages LoRA adapters with HuggingFace’s TRL for model training and vLLM for inference during model evaluation.
• Created novel datasets of 2 million material modifications and 60 thousand multi-turn tool call conversations for band gap and formation energy property improvement tasks. Utilized in Model Evaluation and Supervised Fine-Tuning.
• Implemented a training environment to enable Reinforcement Learning with Symbolic Feedback via Group Relative Policy Optimization (GRPO) and non-sparse, verifiable rewards.
• Wrote optimized GPU Slurm job scripts to orchestrate distributed, multi-gpu post training runs with NVIDIA 80GB H100 GPUs on HPC clusters.

CPSec: Cyber-Physical Systems Security Lab
Apr. 2025 – Present Atlanta, GA

• Investigating internet-exposed maritime infrastructure and vessels by using Censys and Shodan to identify NEMA 0183/2000 protocol signatures.

SymbolicGym: Unified RL Platform for Symbolic Reasoning
Apr. 2025 – June. 2025 Atlanta, GA

• Architected a modular Python framework integrating Reinforcement Learning agents (PPO, DQN, MARL) with formal symbolic reasoning tools (Z3, SymPy, Minisat).
• Designed complex reward functions evaluating syntax, semantics, and domain-certificate verification, enabling agents to iteratively reduce constraints in mathematical and SAT-based search spaces.
• Implemented Graph Neural Network (GNN) state representations to encode logical formula trajectories, significantly improving the agent's interpretability and logic-solving convergence rates.

View the code on GitHub →