Hi I am Nirbhay Sharma!

Mastercard

Data Scientist [Internship]

May2025 - July2025

Multi Teacher GNN To MLP Distillation

• Worked with AI Garage on Multi-Teacher GNN to MLP knowledge distillation for efficient inference
• Trained multiple teacher GNN models like GCN, GAT, SAGE on node and edge prediction, contrastive tasks
• Proposed GMOE2MLP, a graph mixture of expert to MLP distillation method and a novel cluster contrast loss to distill the embeddings from best teacher to MLP

Faaya Astu

ML Engineer [Full Time]

July2023 - July2024

Content Generation, Avatar Generation

• Trained Stable Diffusion ControlNet models on Lineart and Colorbox control on VastAI GPU instance and deployed them on RunPod for more flexibility and control on print generation
• Trained Low Rank Adaptation (LoRA) models using Kohya_SS for custom face and background generation
• Experimented with custom ComfyUI workflows with integrated ControlNet, LoRA, InstantID models
• Containerised ComfyUI with Docker and deployed them as Serverless Endpoints on RunPod and exposed endpoint APIs to AWS Lambda to create APIs for APP using AWS API gateway

Exawizards

AI Engineer [Internship]

June2022 - July2022

Split Neural Network Models

• Worked on Split Neural Network ML paradigm and Splitted Mask-RCNN, FCN_Resnet50, YOLOv5 models for Instance segmentation, segmentation, face detection tasks
• Implemented Autoencoder model for efficient image compression to latent space and setup Pysyft to communicate latents from Jetson Nano to GPU server, preserving data privacy at Jetson Nano

IIT Jodhpur

August2022 - May2023

FedAgPD: Aggregation-Assisted Proxyless Distillation

Supervisor: Dr. Deepak Mishra

• Proposed a novel FL Framework FedAgPD to simultaneously handle model and data heterogeneity
• Leveraged Deep Mutual Learning at Client and Aggregation followed by Gaussian Noise based data free distillation at the Server, eliminating need of proxy dataset or GAN's
• FedAgPD achieved 2x better performance compared to SOTA FL algorithms like FedDF, FedMD, Kt-pfl

IIT Jodhpur

June2021 - May2022

Extremely Lightweight CNN for Chest X-Ray Diagnosis

Supervisor: Dr. Angshuman Paul

• Designed a novel Lightweight CNN model (ExLNet) for the abnormal detection of Chest Radiographs
• Fused Squeeze and Excitation blocks with Depth-wise convolution to create DCISE layer as a component of ExLNet, which outperforms SOTA models like Mobilenet, Shufflenet on NIH, VinBig medical datasets

IIT Jodhpur

August2022 - May2023

Cell Detection and Classification

Supervisor: Dr. Angshuman Paul

• Detected and classified cells data sample into necrotic and apoptotic cells
• Finetuned various SOTA object detectors such as YOLO, SSD, RetinaNet, DeTR
• Achieved remarkable results using DeTR with a Mean Average Precision (MAP) of 40.0