My work asks how AI agents and decision-support systems can remain useful, inspectable, and dependable when they are deployed in environments shaped by uncertainty, evidence requirements, and real operational cost.
Ongoing work is organized around a small number of durable directions in agents, speech pipelines, and deployment rather than a long list of disconnected experiments.
Developing systems that move from raw conversational audio to structured, reviewable outputs while preserving traceability across retrieval, agent steps, and generated conclusions.
Current next step: Refining evaluation slices for transcript quality, retrieval behavior, agent steps, and reviewer trust.
Studying how RAG-based and agentic workflows can support fraud-related conversation analysis without relying on unsupported language-model reasoning.
Current next step: Extending retrieval, tool-use, and answer-grounding evaluation for analyst-facing use.
Exploring how collaborative training setups behave under realistic leakage and privacy assumptions in sensitive AI settings.
Current next step: Comparing attack and defense trade-offs across threat models and deployment assumptions.
Recent English conference papers connect the research agenda with practical deployment, threat modeling, and measurable defense strategy.
Recent conference paper
Presented in English at CISC 2025, this paper analyzes Medusa and CrazyHunter as case studies for AI-era ransomware evolution and connects their tradecraft to a ZTAID-grounded zero-trust defense framework for real operational environments.
Conference: Cryptology and Information Security Conference 2025 (CISC 2025)
Schedule: May 28-29, 2025
Venue: Feng Chia University
Format: Conference Paper · English
Recent conference paper
Presented in English at CISC 2025, this paper proposes Threat Pulse Modeling as a way to convert live cyber threat intelligence into pillar-level ZTAID maturity signals for continuous zero-trust assessment and faster operational response.
Conference: Cryptology and Information Security Conference 2025 (CISC 2025)
Schedule: May 28-29, 2025
Venue: Feng Chia University
Format: Conference Paper · English
Designing AI agents and systems where reliability, evaluation, human review, and traceability are part of the architecture rather than afterthoughts.
Building ASR + LLM + RAG pipelines and agent workflows for conversational analysis, structured extraction, and evidence-aware reasoning over long-form audio and transcripts.
Studying privacy, leakage, adversarial risk, and governance constraints that shape AI systems used in regulated or security-sensitive environments.
I tend to approach AI as a full-system and agent-workflow problem rather than a single-model problem. That means thinking about data quality, tools, retrieval, evaluation, security assumptions, failure analysis, and human review as connected parts of the same design task.
The common thread across the portfolio is simple: build agents and systems that are capable, inspectable, evidence-aware, and realistic about deployment conditions.
I am especially interested in collaborations that value technical depth, careful evaluation, and the realities of deploying AI agents and systems where reliability, reviewability, and governance matter.
Good collaboration fits include research groups, interdisciplinary labs, and technical teams that want to turn strong ideas into inspectable agents, evaluable prototypes, or conference-ready case studies.