Recent advancements in healthcare research have underscored the profound impact of stress on overall well-being and its contribution to the onset of various serious illnesses. Wearable technologies, such as smartwatches and biosensors, have demonstrated strong potential in facilitating more effective stress management. By integrating data from both physical and mental health indicators, these systems can significantly improve medical decision-making, drive innovation in therapeutic research, and enhance our understanding of complex medical conditions. Despite these advancements, conventional Machine Learning techniques often encounter critical challenges, especially in real-time responsiveness, latency, and energy consumption—that limit their deployment on resource-constrained embedded systems. In response, TinyML has emerged as a transformative solution. It enables the execution of deep learning models directly on ultra-low-power devices, allowing for real-time, continuous, and autonomous health monitoring, even in offline or remote environments. In this study, we present a deep learning-based framework incorporating various architectures, including Time-Series Transformers, Autoencoders with Dense Classifiers, Long Short-Term Memory (LSTM) networks, a hybrid CNN + LSTM model, and 1D Convolutional Neural Networks (1D CNN). We emphasize the pivotal role of TinyML in facilitating on-device analysis of Heart Rate Variability (HRV) a vital marker of stress aiming to support personalized health interventions and improve the precision of medical monitoring systems.