This paper presents a custom-trained object detection model based on YOLOv8n, optimized for real-time multiclass detection of animals, vehicles, and people. Leveraging a curated dataset of 2966 instances across 1000 images, the model was fine-tuned with extensive data augmentation techniques, layer freezing, and learning rate scheduling to maximize generalization. Training was conducted for 100 epochs using a batch size of 16 and an image resolution of 640x640. On the test set, the model achieved a precision of 0.753, recall of 0.591, mAP@0.5 of 0.659, and mAP@0.5:0.95 of 0.431, demonstrating a strong balance between accuracy and computational efficiency. The final architecture, with only 3 million parameters, enables deployment on edge devices while maintaining high detection performance. The paper highlights the trade-offs between model complexity, training strategies, and precision targets, providing insights for real-world deployment in constrained environments.