Chengyu "Cathy" Fan

About Me

Hi, I'm Cathy (or go by my legal name: Chengyu, which is a little tricky to pronounce). I graduated from Colby College in May 2026 double majoring in Computer Science (AI concentration) and Mathematical Sciences.

My broad research interest is in building reliable, robust, and secure AI systems. After graduating, I founded Ptero AI Research LLC , an independent AI research and technology consulting company, and began to collaborate with Prof. Zhiyuan Yu on research at the intersection of AI security and mechanistic interpretability. My current work studies how security- and safety-related representations inside large language models change over the course of multi-turn interactions, and how these dynamics may affect existing defense mechanisms such as latent-space detection and activation steering. More broadly, I am interested in understanding how examining models' internal representations can help us identify vulnerabilities and design more reliable and adaptive AI defenses.

My current research direction grew out of my previous work on reliable and multimodal machine learning at Colby. As an undergraduate student, my honors thesis investigated conversation enjoyment through multimodal behavioral and linguistic signals, with an emphasis on modeling dyadic interactions. I also worked with Prof. Tahiya Chowdhury in HUMANE Lab on improving the reliability and robustness of human pose estimation systems by tackling two challenges: annotation quality and edge cases in training data (e.g. occlusion). In addition, I worked with Prof. Stacy Doore on machine learning methods for generating accessible art descriptions in museums for people with visual impairments.

I'm also grateful to have had the opportunity to work on research projects in computer vision, LLM-driven agent simulation, and multimodal interaction for conversation analysis under the guidance of Dr. Amanda Stent , Dr. Michael Yankoski , Prof. Trenton Ford , and Prof. Veronica Romero .

I am currently preparing for the next stage of my research career and am particularly interested in pursuing doctoral research in AI security, adversarial machine learning, mechanistic interpretability, and reliable AI. I am especially excited about research that connects our understanding of models' internal mechanisms with practical questions about robustness, security, and defense.

Research

Watch Demo

Pose Estimation and Responsible Multimodal Interaction

Currently integrating pose estimation techniques with ethical AI principles as part of an ongoing research project in Colby's HUMANE Lab, including proposing an occlusion index for robustness assessment.

Movement Coordination in Dyadic Conversations

This research, combining computer science and psychology, adopted methods such as Cross-recurrence Quantification Analysis (CRQA), Dynamic Time Warping (DTW), and Convergent Cross Mapping (CCM) to quantify human movement coordination.

Demo 1 Demo 2 Demo 3

Comp-Husim: LLM-based agent simulation

At Colby's Davis Institute for AI, I helped build a persistent multi-agent simulation platform (Comp-HuSim: Complex Human Simulation) using large language models to simulate agents with digital personalities.

AI for Environmental Sustainability

Developed a drone image dataset (~1k images) and trained a computer vision model to detect coastal litter in Maine, achieving ~92% accuracy and supporting local recycling efforts.

Publications

Note: * indicates equal contribution