Medical monitoring technologies have gained increasing importance in recent years. Among emerging wearables, in-ear sensing offers a promising alternative to wrist-based devices due to its stable environment and proximity to major arteries, with machine-learning (ML) models showing potential to improve signal analysis performance in this domain, although their design and implementation often lack systematic methodology and reproducibility. This thesis aims to address these gaps by designing an end-to-end in-ear cardiac monitoring system, from custom hardware and dataset collection to the development of a reproducible machine-learning framework for peak detection suitable for embedded deployment. A custom-fit, multi-location in-ear photoplethysmography (PPG) sensing system was developed to collect a multi-activity dataset with a ground-truth electrocardiogram (ECG) reference, enabling systematic evaluation of different Convolutional Neural Network (CNN) architectures for embedded purposes. Results show that signal quality, and thus model performance, strongly depends on sensor placement, with the deep external auditory meatus providing the best signals, followed by the concha. The systematic architecture exploration further revealed consistent design patterns associated with higher accuracy, enabling efficient peak detection with strong ECG correlation. Overall, this work establishes a standardised framework for automatically identifying optimal embedded model architectures for in-ear PPG analysis. Key limitations include the single-subject dataset, computational constraints during model training, and limited final on-device validation.