Deploying artificial intelligence models on embedded systems is a critical challenge in edge computing, where constraints on memory, processing power, and energy efficiency limit the feasibility of traditional deep learning architectures. This study presents a practical approach to implementing a fall detection system on a resource-constrained microcontroller using accelerometer data from the FALLALLD dataset. The proposed methodology includes preprocessing, feature extraction, model training, and conversion to C language using EmLearn to ensure efficient execution on the development board. A comparative analysis of various model configurations explores the trade-offs between accuracy, inference latency, and memory constraints. Results indicate that a 5-s time window, which defines the segment of accelerometer data collected from wrist-worn wearables used for prediction, provides an optimal balance between predictive performance and resource efficiency. However, most trained models exceed the 512 KB flash memory available on the deployment platform. Future work will explore model quantization, alternative architectures, and lightweight AI frameworks to enhance deployment feasibility. This research contributes to the growing field of embedded AI, offering insights into optimizing deep learning models for real-time applications in edge devices.