Working on developing an image engine, triton deployment, and optimization of models • Fine-tuning of 12B parameter model via the Lora technique. Improved inference time (latency) of the model by 11 times via Post-training quantization and quantization-aware training. Fusion of multiple Loras to improve creativity of the model. Made model compatible for serving requests via the Triton server.

Worked on developing algorithms for Identification.• With the help of TensorFlow, Pytorch, FastAPI, remote servers for multi-GPU training (SLURM-based), and algorithms like Transformers, Attention Block, Self-supervised and transfer learning

Detect Image Forgery and Interpretability Methods in Deep Learning under Dr. Arnav Bhavsar's (SCEE Professor at IIT Mandi) guidance. Work on CASIA 1.0 and CASIA 2.0 Dataset. Error-Level-Analysis is done for preprocessing the data. Different VIT-based models were analyzed to localized forgery in an image. Further different Interpretability Methods were tested on the Dataset including Receptive Field and Guided Grad-CAM.

Solved problems involving data structures and algorithms.

Built REST API (MVC Architecture) and built models, routes, and controllers for the same.