Experience 📃

Reseach Internships

Explainability in Graph Neural Networks for Jet Tagging – IIT Kanpur

2026 Annual HEP Conference IITK · arXiv:2604.25885

I worked on developing GNN-based jet-tagging classifiers and combined them with explainable AI tools such as GNNExplainer and Shapley values to examine how the model makes its decisions. This work showed that high-performing models do not need to remain black boxes, and that their predictions can stay physically consistent while revealing the jet substructure features they learn.

Knowledge Distillation in Super-Resolution – Brown University

I worked on applying super-resolution and knowledge distillation techniques to jet classification tasks. My work involved designing CNN-based models for high and low resolution inputs, implementing ESRGAN super-resolution model, and preparing multi-resolution datasets. The goal is to compress models while maintaining performance and improving scalability on large physics datasets.

I worked on a semantic communication framework for visual question answering using MU-DeepSC. I extended the system with Vision Transformer-based encoders and evaluated performance under noisy channel conditions. My work showed improved accuracy when using transformer-based semantic extractors compared to CNN-based baselines.

Course Projects

Privacy in Retrieval-Augmented Generation – CS787: Generative AI

The project focused on evaluating privacy risks in retrieval-augmented generation systems. A privacy evaluation pipeline was developed to study data leakage and mitigation strategies, and multiple defenses such as PII sanitization, summarization, and prompt engineering were benchmarked. The project also included membership inference audits using S-RAG to analyze privacy–utility trade-offs and implicit privacy risks in practical RAG deployments.

Unsupervised Domain Adaptation – CS771: Introduction to Machine Learning

This project focused on learning under domain shift without labeled target data. I studied methods such as pseudo-labeling, RandMix, and prototype-based alignment and evaluated their effectiveness in improving cross-domain generalization. The project emphasized practical challenges in deploying ML systems across changing data distributions.

Predictive Modeling and Representation Learning – CS771: Introduction to Machine Learning

This project entailed building binary classification models across three datasets using approaches such as n-grams, XGBoost, and deep neural networks. We experimented with multiple feature representations and performed systematic hyperparameter tuning using GridSearchCV. The models achieved strong performance across datasets, highlighting the impact of representation choice.

Pahal Patel