Sudden cardiac arrest, often caused by ventricular fibrillation (VF) and ventricular tachycardia (VT), is a leading cause of death worldwide. Timely detection and treatment of these shockable rhythms using Automated External Defibrillators (AEDs) are crucial. This work presents a Deep Learning (DL)-based system for real-time detection of shockable rhythms using a single-lead electrocardiogram (ECG) and limited computing resources. Following the principles of Frugal AI, we focus on developing an efficient, lightweight model that balances high classification performance with minimal hardware requirements. We achieve this by performing Bayesian hyperparameter optimization on the ResNet architecture and applying 8-bit quantization. We also benchmark the model on several STMicroelectronics development boards, assessing performance, the effect of quantization, and resource requirements, with the aim of ensuring computational efficiency and low latency without compromising detection accuracy. Our best results were able to detect shockable rhythms in under 3 s with 88.9% sensitivity and 99.6% specificity with 0.2 s of inference time using an STM32 NUCLEO board. Our approach adheres to international AED guidelines and standards, demonstrating its practicality for the healthcare market.