I am a PhD student at WPI studying data science and robustness in AI. My current work focuses on making LLMs more robust against different perturbations and prompt styles, and improving uncertainty estimation for language models. My published work focuses on assessing the robustness of U.S. strategic highways and understanding how natural disasters, such as wildfires, impact critical transportation networks.
Sukhwan Chung, Daniel Sardak, Maksim Kitsak, Andrew Jin, Igor Linkov
Military logistics rely heavily on public infrastructure, such as highways and railways, to transport troops, equipment, and supplies, linking critical installations through the Department of Defense’s Strategic Highway Network and Strategic Rail Corridor Network. However, these networks are vulnerable to disruptions that can jeopardize operational readiness, particularly in contested environments where adversaries employ non-traditional threats to disrupt logistics, even within the homeland. This paper presents a contested logistics model that utilizes network science and Geographic Information System (GIS) to evaluate the robustness and resilience of strategic transportation networks under various disruption scenarios. By integrating GIS data to model logistics networks, simulating disruptions, and quantifying their impacts, we identified vulnerabilities in US power projection routes and assessed the resilience and robustness of highways and railways. Our findings reveal that highways are more resilient than railways, with greater capacity to absorb targeted disruptions.. These findings underscore the importance of prioritizing investments in highway infrastructure and reinforcing vulnerable road and rail segments, particularly in high-risk regions, to enhance the resilience of military logistics and maintain operational effectiveness in contested conditions.
Sukhwan Chung, Daniel Sardak, Jeffrey Cegan, Igor Linkov
Network science is a powerful tool for analyzing transportation networks, offering insights into their structures and enabling the quantification of resilience and robustness. Understanding the underlying structures of transportation networks is crucial for effective infrastructure planning and maintenance. In military contexts, network science is valuable for analyzing logistics networks, critical for the movement and supply of troops and equipment. The U.S. Army’s logistical success, particularly in the “fort-to-port” phase, relies heavily on the Strategic Highway Network (STRAHNET) in the U.S., which is a system of public highways crucial for military deployments. However, the shared nature of these networks with civilian users introduces unique challenges, including vulnerabilities to cyberattacks and physical sabotage, which is highlighted by the concept of contested logistics. This paper proposes a method that uses network science and geographic information systems (GIS) to assess the robustness and resilience of transportation networks, specifically applied to military logistics. Our findings indicate that, while the STRAHNET is robust against targeted disruptions, it is more resilient to random disruptions.
Stable Diffusion, LLM Bias, Text-to-Image Models | Jan 2025 - May 2025
Analyzed bias of several text-to-image models using a combination of no-reference automatic image quality metrics and human-annotated scores. We analyzed images of objects in different continents (for example wedding dress in Africa). Our results showed a disparity between the automatic and human metrics - while the human metrics indicated that typically more underrepresented regions tended to feature more stereotypical images, the automatic metrics gave them higher quality scores.
RAG, LLM, Haystack | Jan 2025 - May 2025
Developed a RAG-powered immigration form filler using Mistral and Haystack. Features an end-to-end RAG pipeline with a chatbot interface for document upload and user information collection.
Python, Reinforcement Learning | Aug - Dec 2024
Implemented a Rainbow DQN agent with reward shaping techniques. Trained PPO and A2C agents achieving scores of 2,200+.
LLM, SQL, Python | Aug - Dec 2023
Created an LLM-powered database assistant using GPT-3.5-turbo that converts natural language queries into SQL. Includes a chatbot interface for database interaction.