ardiovascular diseases remain a leading cause of death worldwide, and arrhythmia, an irregularity in heart rhythm, demands timely detection. Conventional ECG monitoring is bulky, hospital-bound and computationally heavy, which limits continuous monitoring in everyday settings. This paper presents an IoT-based arrhythmia detection system that combines a lowcost wearable ECG front end with a lightweight deep learning model running directly on the sensor node. Biopotentials are acquired with ECG electrodes, conditioned by an AD8232 analog front end, digitised and processed by an ESP32 microcontroller, and classified by a one-dimensional convolutional neural network (DL-LAC) into five heartbeat classes: Normal (N), Supraventricular (S), Ventricular (V), Fusion (F) and Paced (P). Because the network learns features directly from raw single-lead ECG, no manual feature engineering is required. The model, trained on the MIT-BIH Arrhythmia Database, reaches approximately 96.2% overall accuracy, with on-device inference latency of about 80 ms per window and a memory footprint below 500 KB. Duty cycling reduces power consumption by about 40%, extending battery life from four to over six hours. The results indicate that on-device deep learning is a practical route to continuous, low-latency and privacy-preserving cardiac monitoring.