The field of robot-aided rehabilitation has recently witnessed remarkable advancements, with the emergence of robotic devices that incorporate visual, auditory, and multimodal sensory stimulation to enhance rehabilitation for both musculoskeletal and neuromuscular patients. Additionally, combining data-driven methods and Artificial Intelligence algorithms facilitates increasingly customized and effective approaches to physical rehabilitation. This paper introduces the AI-CARE system, a robotic architecture for upper limb motor rehabilitation that integrates a deep reinforcement learning method for detecting physical and cognitive stress. The system generates customized interventions ranging from the modulation of the physical assistance provided by the robot to multimodal stimulation taking advantage of wearable multimodal monitoring of the user. In its preliminary phases, this research investigates whether physical and cognitive stress can be detected from physiological data collected via wearable sensors, if different physiological responses can be measured through robot physical assistance modulation or multimodal stimulation, and whether real-time indicators of individual effort as a function of task demand can be extracted from the electroencephalographic signal. Early results suggest that the stress detection module integrated into the system outperforms a baseline model in both classification accuracy and response time. Furthermore, initial evaluations of the selected interaction modalities show the potential to induce varied physiological responses in both the autonomic and central nervous systems, paving the way to further developing customized adaptation strategies.