Hazelnuts are a valuable agricultural product with significant economic and nutritional importance, particularly in countries like Turkey, which leads global production. Ensuring the quality of hazelnuts during post-harvest processing is essential to maintain market standards and consumer satisfaction. Traditional inspection methods are often labor-intensive, inconsistent, and inefficient, highlighting the need for automated and intelligent systems. In this study, we propose a lightweight deep learning model implemented on the Edge Impulse platform to classify hazelnuts into three categories: Hole, Crack, and Good Quality. The dataset was labeled using the Edge Impulse image annotation tool and divided into training (81%) and testing (19%) subsets. The MobileNetV2 0.35 model was employed through transfer learning, achieving a validation accuracy of 97.2% and a test accuracy of 93.36%. The model demonstrated efficient real-time performance with a processing speed of 827 ms per image and low peak RAM usage of 119.4 KB, making it highly suitable for deployment on edge devices in industrial settings. Compared to previous studies that relied on complex or resource-heavy architectures, our model balances high accuracy with minimal computational requirements. The findings validate the potential of lightweight deep learning models in automating hazelnut quality assessment and offer a scalable solution for small and medium-sized processing facilities. This study represents a step toward smart agriculture through embedded AI-driven quality control.