I am a postdoctoral researcher at the CMU Safe AI Lab, where I work on blockchain agent validation — building agentic AI systems that audit and verify on-chain transactions. My broader research studies the safety and security of machine learning systems, spanning multimodal models, federated learning, and agentic AI. I completed my PhD at HKUST, advised by Prof. Songze Li and Prof. Yangqiu Song.
My research studies how modern AI systems fail — and how to make them safer. I uncover vulnerabilities in vision-language models and federated learning systems, design privacy-preserving and robust learning methods, and build agentic AI systems for high-stakes domains such as blockchain security and automated DeFi auditing.
Backdoors, memorization, hidden behaviors, evaluation, and safety risks in multimodal foundation models.
Privacy attacks, reconstruction risks, adversarial robustness, and secure vertical federated learning.
Auditable LLM agents, dynamic validation graphs, and failure attribution for security-critical workflows.
Building reliable ML pipelines that connect algorithmic guarantees with practical deployment constraints.
Building agentic AI systems for blockchain agent validation — integrating graph-of-thought reasoning with on-chain validation pipelines for automated DeFi auditing. Also developing RL-based autonomous agents with intrinsic-motivation skill discovery.
Built an auditable LLM validation agent for DeFi security with node-level failure attribution. Designed a graph-based validation pipeline evaluated on 18K+ transactions across three protocols.
Developed Hidden Ads, a behavior-triggered semantic backdoor attack for advertisement injection in vision-language models, revealing emerging multimodal threat vectors.
I am interested in research scientist, applied scientist, and ML/security roles where I can build trustworthy AI systems — especially at the intersection of foundation models, autonomous agents, privacy, and security-critical applications.