I’m a Ph.D. researcher at Korea University’s NLP&AI Lab, specializing in Mechanistic Interpretability of Large Language Models (LLMs). My work focuses on reverse-engineering the inner workings of LLMs to uncover their decision-making processes, enhance transparency, and ensure alignment with human values. I am passionate about advancing AI research responsibly by bridging theoretical insights with practical applications.

• Mechanistic Interpretability
  Reverse-engineering LLMs to understand their internal circuits, algorithms, and emergent behaviors. My work includes:

  • Developing sparse autoencoders to isolate interpretable features in transformer architectures
  • Probing causal relationships between model components and specific capabilities
  • Mapping information pathways in large-scale models to improve transparency and reliability

• AI Safety
  Ensuring the safe and ethical deployment of AI systems by focusing on:

  • Automated attack detection (harmfulness/bias detection) using red-teaming techniques
  • Mitigating biases in LLMs through interpretability-driven methods
  • Designing scalable frameworks for aligning AI behavior with human values

• Mechanistic Anomaly Detection
  Identifying unexpected or harmful behaviors in LLMs by analyzing their internal mechanisms. Key contributions include:

  • Developing tools to trace causal pathways for diagnosing anomalous outputs
  • Designing self-monitoring models for real-world deployment scenarios
  • Improving robustness under adversarial or high-stakes conditions

• Reasoning Models & Agent Systems
  Investigating how LLMs reason and interact as agents to perform complex tasks reliably. My research focuses on:

  • Multi-step reasoning processes within transformer-based architectures
  • Building RAG agent systems for dynamic knowledge retrieval and integration
  • Exploring compositionality in neural networks for structured reasoning

• Retrieval-Augmented Generation (RAG)
  Enhancing generative models with retrieval mechanisms to improve factual accuracy and groundedness. My work includes:

  • Designing retrieval pipelines optimized for domain-specific applications
  • Reducing hallucinations by embedding retrieval mechanisms into transformer workflows
  • Improving factual consistency in generative outputs through hybrid architectures

• Cognitive Alignment & Ethical AI Design
  Ensuring that AI systems operate consistently with ethical principles and human intentions. Contributions include:

  • Embedding ethical guidelines into model training processes through scalable fine-tuning methods
  • Leveraging interpretability tools to monitor alignment over time
  • Collaborating across disciplines to design responsible AI frameworks

• Mechanistic Interpretability: Developing novel techniques to analyze the internal structures and decision-making pathways of LLMs.
• AI Safety & Bias Mitigation: Creating robust frameworks to ensure ethical and safe deployment of advanced AI systems.
• RAG & Agent Systems: Designing retrieval-augmented generation pipelines and agent-based systems for dynamic decision-making tasks.

• LlamaIndex: Efficient data indexing and retrieval for LLM-driven applications.
• Haystack: End-to-end NLP framework for building search systems and conversational agents.
• LangFlow: Tool for orchestrating complex workflows in LangChain-based applications.
• Helicone: Observability platform tailored for monitoring LLM-powered applications.
• Gemini (Google): Advanced multimodal language models optimized for enterprise-level tasks.

As a Ph.D. researcher at the NLP&AI Lab under Dr. Heui-Seok Lim, I have contributed to government and industry-funded projects, including collaborations with the Ministry of Food and Drug Safety and KT Gen AI Lab. Key projects include:

• Developing a novel knowledge editing method for domain-specific applications without retraining models.
• Designing automatic attack detection frameworks (harmfulness/bias detection) using red-teaming techniques.
• Creating advanced RAG agent systems capable of dynamic knowledge retrieval for real-time decision-making tasks.

During my B.S. in Computer Science, I worked under Dr. Wanwan Li on augmented reality systems, focusing on automatic room mapping using SLAM algorithms. Additionally, I collaborated with Dr. Edwin Michael to develop agent-based models for pandemic simulations in Hillsborough County, contributing to public health planning during COVID-19.