Foundation models are rapidly moving from controlled research into critical real-world infrastructure in robotics, clinical decision support, industrial inspection, and edge devices, where environments shift, failures carry consequences, and compute is constrained. Yet a deep gap remains between impressive AI demonstrations and safe, robust deployment. Closing this gap is not merely a technical challenge; it is essential for AI to realise its transformative potential.

Real-world deployment exposes three fundamental limitations in current AI systems: limited adaptiveness (models remain static and require costly retraining), insufficient trustworthiness (poor calibration and overconfident predictions in high-stakes settings), and resource inefficiency (reliance on monolithic architectures).

My research addresses these challenges through three directions: Adaptive AI, Trustworthy AI, and Resource-Efficient AI, unified through the concept of a Modular AI System (MAS). Rather than treating AI as monolithic, MAS reconceptualises it as interacting, independently configurable components built around a shared core. This framework enables fine-grained adaptation, well-calibrated behaviour, and efficient deployment within a unified architecture.