- Designed a novel architecture for clustering short text documents using Large Language Models (LLMs).- Built a prototype clustering application in Python using pretrained GPT-3.5 and GPT-4 LLMs.- Researched various prompt engineering approaches to maximize accuracy and reduce hallucination.

- Research towards differentially private (DP) multi-document summarization models for user feedback cluster summarization using large language models such as BART and T5.- Contributed to the opensource library microsoft/dp-transformers to make T5 model training compatible with Opacus, by implementing sample level gradients in PyTorch for non-standard hugging face library model layers- Introduced a novel application and analysis of differential privacy for a multi-document task

Collaborated with MS Research and NLP experts in research towards using differential privacy (DP) and synthetic data generation for overcoming privacy concerns for customer feedback text data- Generated high utility and privacy-preserving labeled synthetic data using transformer-based GPT2 model- Designed a novel approach in PySpark for empirical estimation of true training data leakage by releasing synthetic text data- Trained DP privacy-preserving BERT models with production level accuracy for the task of feedback’s topic classification- Explored the use of DP n-gram extraction method for privacy-preserving text data release

Synthetic Data Generation with Privacy Protection- Research towards a generic solution with theoretical guarantees for utility and privacy of synthetic data generation for tabular, image, and multimodal data- Developing new architectures leveraging existing deep learning networks together with adversarial networks and autoencoders to synthesize data (Python, PyTorch, pandas, scikit-learn)

Mathematics of Data Science - Spring 2022Probability - Winter 2021Calculus A - Fall 2020Multi-variable Calculus - Spring 2020Applied Linear Algebra - Winter 2020Multi-variable Calculus - Fall 2019

Ticketing System for User support:- Designed and implemented an intelligent ‘Ticketing System’ with Knowledge Base and self-service functionalities- Developed end-to-end application from UI to backend – Angular based JavaScript web application, Camuda based workflow to manage lifecycle of a ticket, Java based RESTful API architecture- Mentored 2 summer interns to develop a microservice – historic data maintenance in elastic search, metrics collection via Kafka, data visualization – enabling operation managers to analyze the tickets and monitor the performance of support teams- Led the onboarding efforts resulting in onboarding of 12 teams with 56 applications by collaborating with stakeholders

Kiosk application for Secured access to Sensitive Production Environment:- Eliminated 73% of direct production access by developing an intermediate application to run secured shell scripts- Developed a key functionality to enable end-users to raise multiple consecutive shell executions in one request, increasing the platform use by about 500%- Worked with UX team to revamp the app to enhance customer satisfaction and implement highly demanded missing functionalities. Also included an upgrade of front-end from AngularJS to Angular 6- Became a Subject Matter Expert for technical and business aspects of the applicationAutomated Data Migration Tool from Legacy to Kiosk Application:- Worked in collaboration with Tech-Risk and Production access control team on a time and risk-sensitive automated tool (built with Java, Angular, Camuda) to migrate all firmwide operation support shell scripts from legacy to new Kiosk application- Conducted multiple regional and application specific trainings smoothing the onboarding of about 3000 support team members

Proposed and implemented in Java, a model for streaming videos to users in remote villages, with low internet connectivity and low memory end user devices.