Biography
Hi, I am a first year PhD student in the Department of Computer Science at the University of Texas at Austin, advised by Prof. Işıl Dillig and Prof. Greg Durrett. I work on secure code transpilation using neurosymbolic techniques.
I have, previously, been a Pre-doctoral Research Fellow at Microsoft, working with the PROSE (Program Synthesis) team where I worked on the Copilot experience for data wrangling in Fabric. I received my bachelor’s degree in Information Technology from Veermata Jijabai Technological Institute (V.J.T.I.), Mumbai, India.
I have had the good fortune to work with Dr. Ashish Tiwari, Dr. Sumit Gulwani, Dr. Vu Le, Dr. Gust Verbruggen, Dr. Sandeep Udmale, and Dr. Vijay Sambhe.
Outside of work, you can find me running, playing the guitar, and listening to songs.
Download my CV .
- AI4Code
- Artificial Intelligence
- Information Retrieval
- Program Synthesis
Ph.D. in Computer Science, started 2024.
University of Texas at Austin.
B.Tech. in Information Technology, 2017-2021
Veermata Jijabai Technological Institute, India.
Recent News
[10/10/2024] I am on the PC for ICLR, 2025.
[01/10/2024] I am on the PC for the Table Representation Learning Workshop at NeurIPS ‘24.
[22/05/2024] I am serving on the PC for ASE-Industry Track 2024.
[05/05/2024] I will be on the PC for OOPSLA ‘24.
[15/04/2024] 🏆 Our paper “Semantically Aligned Question and Code Generation for Automated Insight Generation” is selected as the Best Paper at LLM4Code at ICSE ‘24.
Recent Publications
Experience
- Conceptualized and built the natural language to code feature for the Power Query M language, used for wrangling tables in Excel, Fabric and PowerBI.
- Collaborated towards building the Copilot experience as a part of the Power Query experience in Fabric and Excel.
- Devised two state-of-the-art strategies TSTR (EMNLP-Findings 2023) and COOPER (Under submission to ASE 2024) for optimal dynamic prompt construction aiding in-context learning for natural language to code tasks.
- Developed Alternate Task Technique (ATT) (Under submission), a generalized framework to post-process LLM outputs using alternate tasks that improved performance on low resource languages, like Power Query M, by 13%.
- Developed Adapted Dense Retrieval (ADDER) (Under submission) framework for Information Retrieval tasks using dense embedding for efficient code retrieval in low-resource settings.
- Collaborated with Microsoft Edge team for web-based data extraction tasks to improve product purchasing experience.
- Successfully automated invoice data extraction tasks for the Microsoft IDC Finance team to improve productivity.
- Employed techniques to combat low-resource name entity recognition tasks by employing ML and program synthesis techniques.
- Devised Landmark-based Robust Synthesis (LRSyn), a state-of-the-art interpretable data extraction framework, robust to version changes in data.
- Spearheaded the clustering and landmark detection tasks during the development of LRSyn, and developed a novel fingerprinting technique for images.
- Successfully published our research paper titled “Landmarks and Regions: A Robust Approach to Data Extraction” at the Conference on Programming Languages Design and Implementation 2022, San Diego.
- Collaborated with 30 changemakers to develop a war-crime detection tool using social media channels.
- Fine-tuned a distil-RoBERTa model for binary classification of war crimes
- Spearheaded the development of a novel two stage prediction pipeline for multi-label classification of warcrimes.
- Worked with the On-Device AI team to improve system performance using Reinforcement Learning.
- Built a State-Of-The-Art Multi-Agent Deep Q-network leveraging prioritized experience replay(PER) and time-bound dynamic reward functions
- Designed a landmark agent simulation environment to show proof of concept.
- Coordinated with the product development team to build an ERP application for manufacturing and service-based industries.
- Employed JavaFX for the development of the application and MySQL for database management.
- Followed an agile based product development life cycle with constant interaction with key product owners.
Accomplishments
