• Enhanced YOLO V7's performance by developing custom models incorporating advanced neural network modules such as CBAM and ConvNeXt, alongside transformer-based techniques like RTM-Seg. -This innovation increased model accuracy from a MAP50 of 89.4% to 97.8%. • Implemented ECAnet, TransNext, SENet, CSPResNext, EELAN,SPP,etc from research papers. Boosted classifier accuracy from 85.2% to 94.27% on test dataset. • Implemented YOLOACT++, YoloV9, RTDETR, RTM-det seg, and CenterNet++ for segmentation tasks, achieving a mask-map:50 of 0.869 respectively.

• Automated ML models retraining and updating them on edge devices using Nvidia's DeepStream. • Utilized unsupervised learning techniques to reduce data annotation time by 30%, employed self-supervised learning to boost classification accuracy by 20%.• Integrated Conditional GANs to enhance the robustness and increase the variance of our models to meet diverse operational demands. • Designed and implemented a robust cloud infrastructure for real-time image processing, leveraging AWS services like ECR, AWS Batch, MLflow on EC2, and orchestrated through Lambda functions

• Designed and developed a Cheque Processing System leveraging Deep Learning for Handwritten and Machine-Printed text recognition. • Implemented CRAFT (Character Region Awareness For Text Detection) for text detection and TROCR for text recognition, achieving accurate results.