Hi, I'm Chandreyi (Zini) Chakraborty
BS/MS CS @ Georgia Tech | AI Interpretability Research
I earned my Masters and Bachelors in Computer Science from Georgia Tech, both with machine learning concentrations. My research focuses on uncovering LLM inner workings through mechanistic interpretability, and leveraging those insights to improve model capability and efficiency.
I second-authored a paper accepted to COLM 2026 that attributes social reasoning to pre-training data at the corpus level using influence functions and machine unlearning. This was in collaboration with EleutherAI as a member of the Entertainment Intelligence and Human-Centered AI Lab under Dr. Mark Riedl. I also published on how LLMs understand suspense in stories (COLM 2025).
I was a Machine Learning Teaching Assistant under Dr. Mahdi Roozbahani, where I developed a Socratic LLM agent that supports TAs in responding to student questions on EdStem with LoRA fine-tuning and dynamic RAG databases.
On the efficiency side, I have worked with different LLM architectures and GPU configurations for optimal compute and accuracy. I co-first-authored a paper on the quantization robustness of diffusion LLMs against autoregressive LLMs.
Previously, I worked at Bloomberg on the Terminal, building an efficient portfolio aggregation service, and before that at Deloitte, where I worked with the Tennessee state government to enhance their Eligibility Benefits Management Portal and integrate SummerEBT Appeals.
Publications
- 2026 Where Does Social Reasoning Come From? Capability Provenance in Language Models Conference on Language Modeling (COLM) Also at COLM workshops: Sci-FM and Social Sim'26 arXiv GitHub Website HF
- 2026
- 2025 Do Language Models Agree with Human Perceptions of Suspense in Stories? Conference on Language Modeling (COLM) arXiv GitHub
- 2025 Diffusion Model for Generating Inorganic Materials with Targeted Density of States Georgia Institute of Technology Repository
Contributions
- 2026
Added support and tests for mechanistically interpreting Google’s T5Gemma2 multimodal model.
GitHub - 2024
Co-implemented an SDK, email/Amazon S3 file services, and user onboarding for faster app deployment.
GitHub - 2024
Co-built a donation and event coordination platform for the non-profit, giving chapter organizers a single place to manage retreats for teens fighting cancer.
GitHub - 2026