About

I’m a Ph.D. researcher at Nanyang Technological University working on trustworthy AI for robotic and autonomous systems. I care about making learning-based systems reliable in the real world: exposing failure modes, detecting them at runtime, and designing policies that handle uncertainty safely.

• robust perception & decision-making
• uncertainty estimation and monitoring
• safety evaluation, testing, and simulation
• ML systems that must work under real-world constraints

I am actively seeking AI or robotics roles, with a focus on autonomous systems and trustworthy AI. I am open to both industry and research opportunities—please feel free to contact me if my experience is a good fit.

Research focus

My work sits at the intersection of:

• Adversarial / stress testing: generate rare, hard cases to reveal weaknesses
• Online monitoring: detect distribution shift, uncertainty, and anomalies at runtime
• Decision-making under uncertainty: safer planning/control when information is imperfect

Design philosophy: failure-aware AI

I treat “trustworthy” as a lifecycle: find failures → measure them → reduce them → monitor them in deployment. This mindset applies to many robotics settings (mobile robots, manipulation, drones, and other safety-critical autonomy), not just one domain.

Experience (engineering mindset)

Before my Ph.D., I worked as a software engineer at ZTE, developing high-frequency algorithms (up to 240kHz) for 5G systems. That experience pushed me toward building systems that are deployable, efficient, and testable—not just good in a paper.

An Approach to Trustworthy AI

My research is centered on adversarial design. Trustworthy AI is not achieved by average-case performance, but by continuously exposing and managing failure modes. Illustrated here using autonomous driving as a safety-critical testbed, but applicable to broader AI and robotic systems.

-Adversary actively generates rare, risky, and adversarial scenarios to expose system weaknesses, using passive and active testing as well as objective and subjective evaluations.

-Adaptor continuously improves perception, decision-making, and control to mitigate the failures revealed by adversarial exposure.

-Collaborator (Human-in-the-loop) provides human assistance when uncertainty, ambiguity, or risk exceeds the autonomous system’s capability.