Toward AI safety as a capability maintained throughout operation,
not merely a property specified at design time.
My research focuses on AI safety for adaptive and autonomous systems that learn and operate in changing environments. I study how safety constraints can remain valid as operating conditions evolve, how emerging risks can be anticipated before they lead to unsafe behavior, and how runtime safety mechanisms can intervene when reliable adaptation is not possible.
Forecasting environmental context and adapting safety constraints before changing conditions lead to unsafe behavior. The CoLLAs evaluation reduced collision rate by 87.2% on the main benchmark and 92–97% across three held-out road layouts.
View paper →A modular framework combining context-based, adaptation-speed, and soft-to-hard safety mechanisms for stationary, seen-nonstationary, and unseen-nonstationary conditions.
View paper →Projects cumulative safety budgets into enforceable action-level constraints during operation; reduced average collision rate by 91.6% versus an unshielded DQN baseline across four driving tasks.
View paper →Ultimately, I am interested in making
adaptation itself a subject of safety assurance.