Diabetic Retinopathy (DR), a serious diabetic complication characterized by retinal blood vessel damage, represents a leading cause of preventable blindness worldwide. Early detection is critical, but manual screening by ophthalmologists is labor-intensive and unscalable, failing to meet rising demand. This paper proposes FundusInsight, an automated framework for DR detection and grading using deep learning on fundus images. This paper depicts a refined ResNet50 network that can automatically assess the severity of diabetic retinopathy under five clinically relevant categories namely: No DR, Mild, Moderate, Severe, and Proliferative. The proposed model is comparatively tested on the basis of popular transfer learning networks, i.e., VGG16, ResNet152, and MobileNet V2, to determine its effectiveness on a publicly available dataset consisting of 3,662 retinal fundus images. It is tested and shown that the ResNet50 model has higher performance with a training accuracy of 98.71% and a test accuracy of 73.33%. The model also demonstrates high levels of discriminative potential in identifying non-pathological cases with an F1-score of No DR with a value of 0.96. The proposed FundusInsight framework can be used to decrease clinical workload, aid ophthalmologists to detect the disease at the early stages, and deploy the proposed system to systems of teleophthalmology on a scale, which can further improve patient treatment and results.