Fall detection among elderly is a vital public health issue and requires the development of systems which not only have high accuracy but are also resilient to false alarms and adaptable to different types of behavior. The existing systems, however, lack comprehensive incorporation of three crucial aspects: multi-scale temporal feature learning, global context awareness and adaptive decision making required for practical application. In this work we propose TCN-Transformer-SAC a hybrid deep reinforcement learning architecture that incorporates Temporal Convolutional Network (TCN), Transformer encoder and Soft Actor-Critic agent. The TCN component effectively learns multi-scale local temporal dependencies from inertial sensor signal streams, while the Transformer component uses self-attention to capture global contextual dependency over the whole activity sequence. Finally, the SAC algorithm exploits the learned spatiotemporal features to develop an adaptive decision-making policy that optimally balances the conflicting objectives of maximizing sensitivity and minimizing false alarm rate. When evaluated on public UP-FALL dataset, our system reaches state-of-the-art results, with accuracy 99.0%, precision 100%, recall 97%, and F1-score of 99%. Importantly, the developed framework also shows superior resilience, reaching false positive rate of just 1.2% on difficult cases of fall-like Activities of Daily Living (ADLs). In this case, if the algorithm is deployed on the embedded system NVIDIA Jetson Nano, then it provides real-time inference with latency of 9.3ms and low energy consumption (<5W). This indicates that there is tremendous possibility of leveraging the benefits of temporal convolutions, attention, and reinforcement learning in order to develop a solution for proactive elderly care.