Atrial fibrillation (AF) detection is pivotal for stroke prevention, yet the deployment of robust deep learning models on resource-constrained wearable devices remains a formidable challenge due to excessive computational demands. This paper presents an automated, hardware-aware design pipeline for ultra-lightweight AF classification, driven by Large Language Model-based Neural Architecture Search (LLM-NAS). By translating hardware constraints into structured linguistic priors, we leverage the reasoning capabilities of LLMs to discover an optimized convolutional neural network architecture that synergizes time-frequency dual-branch feature extraction, depthwise separable convolutions, and channel attention mechanisms. To further bridge the gap between algorithmic complexity and embedded efficiency, the discovered model undergoes a two-stage compression suite involving structured pruning and quantization-aware training. Experimental results on the CPSC2021 dataset demonstrate that the resulting model achieves a high F1-score of 0.9674 with only 7.93K parameters and a minimal memory footprint of 7.7 KB—a significant reduction compared to existing state-of-the-art models. Furthermore, we implemented a complete prototype system on an STM32F767 microcontroller, achieving a single-inference latency of 306.20 ms and an incremental power consumption of 0.18 W. This end-to-end validation confirms the feasibility of our LLM-driven methodology for real-time, high-fidelity AF monitoring in next-generation medical IoT devices.