This paper presents a lightweight, real-time fruit object detection and grading system deployed on embedded hardware, specifically Raspberry Pi 4 and Raspberry Pi 5. The system accurately detects and classifies multiple fruit types and quality grades from images to support automated sorting and quality assessment. The pipeline integrates image acquisition, preprocessing, and model inference, leveraging models such as EfficientNet and TinyViT for high accuracy and low latency in resource-constrained environments. Experimental results demonstrate an average inference latency of approximately 150 ms per image with detection accuracy exceeding 92% across fruit classes, while power consumption remains below 10 W, enabling off-grid deployment. Future work will consider expanded datasets, multimodal sensors, and further model optimization for enhanced edge performance.