Diabetic-Retinopathy is a serious diabetes complication of the retina and if not detected early can cause permanent vision loss. The conventional manual detection of Retinopathy through affected eye image analysis by ophthalmologists is time consuming, manual and resource intensive. Deep learning and specifically Convolutional Neural Networks have shown great success in medical image analysis in recent years and can provide efficient and automated solution for DR detection. The existing methods suffer from performance degradation due to class imbalance, computational complexity and long training times. To overcome these issues this research proposes a new DR classification model that combines CNN with embeddings and residual blocks to improve feature extraction and classification accuracy. Data augmentation techniques like flipping, scaling, rotation and contrast modification are used to reduce class imbalance and improve model generalizability. Preprocessing techniques like noise removal and contrast enhancement are also used to improve the quality of retinal fundus images. Experimental results on benchmark dataset show that this proposed model is better in accuracy, sensitivity and specificity than the state of the art. This work uses deep learning and embeddings for automatic DR diagnosis and can contribute to early detection and on-time intervention to prevent visual impairment for diabetic patients.