Sabarna Choudhury

Dynamic software engineer with over 5 years of experience in full-stack development (Angular, .NET Core) at Cognizant, now pivoting toward GenAI and ML engineering. Graduated top of the class with an M.Tech. In Data and Computational Sciences from IIT Jodhpur, earning the silver medal for academic excellence. Co-authored a research paper on LLM fine-tuning for self-harm detection, accepted at ACL 2025, and am actively working on GAN-based synthetic text generation projects leveraging LLaMA 3.1. Passionate about bridging software engineering and GenAI solutions for scalable, production-ready applications.

• Graduated top of class and awarded the Silver Medal for Academic Excellence during M.Tech in Data and Computational Sciences at IIT Jodhpur in 2025
• Co-authored research paper accepted at ACL 2025 on fine-tuning LLMs for self-harm detection using emoji interpretations
• Awarded WorkSight Star Developer for exceptional contributions and innovative solutions – Q1 2022
• Recognized as the best developer in the WorkSight MDP team under the quarterly R&R MLEU awards – Q2 2022
• Achieved Best Developer title in the WorkSight MDP team for outstanding performance and quality delivery, Q1 2023

• GitHub Foundations Certification – Valid until 14th March 2028
• Cognizant Internal – GitHub Copilot Generic Knowledge Based Assessment [201 Intermediate] (Mar 2024)
• Udemy – GitHub Copilot: Use AI to Write Code for You! (Completed Mar 2024)
• AWS Partner – Cloud Practitioner Essentials (2022)
• Udemy – Understanding TypeScript (2021 Edition)
• Udemy – ES6 JavaScript: The Complete Developer’s Guide

1. The Devil Is in the Details: Enhancing LLMs for Self-Harm Detection, 2024 – 2025

As part of my M.Tech at IIT Jodhpur, contributed to a research project accepted at ACL 2025. Worked on designing the SHINES dataset to capture nuanced signals of self-harm intent in social media posts, including emoji interpretation through the CESM-100 matrix. Implemented data curation, annotation pipelines, and experimental evaluations using Hugging Face Transformers. Focused on fine-tuning large language models for both classification and span extraction tasks, achieving significant improvements over baseline methods in detecting self-harm risk.

2. Synthetic Data Generation Framework using LLM-GAN, 2024 – Ongoing

Working on an innovative framework combining LLMs (e.g., LLaMA 3.1) and GAN principles to generate high-quality, emotion-aligned synthetic text for low-resource NLP domains. Contributing to blueprint-based prompt engineering, discriminator feedback loops, and experiments on datasets like CEASE and Dreaddit, achieving SOTA results in emotion and stress detection tasks.