Biography
I am a graduate researcher in Computer Science Engineering, focusing on NLP, Interpretability, Social Computing, and Safety & Alignment in Large Language Models. My research builds pipelines that analyze biases and causal mechanisms within transformer models.
I am currently a Visiting Undergraduate Research Fellow at Harvard University and a Research Assistant Intern at Northwestern University. I recently collaborated with the University of Oxford on mechanistic interpretability and AI safety, and previously worked with the University of Hamburg (HCDS Lab) and NIT Agartala.
Interests
- Interpretability
- NLP
- Social Computing
- Safety & Alignment
Education
-
B.E. in Computer Science Engineering
Visvesvaraya Technological University (VTU), India
Nov 2022 – Aug 2026 · 79% (First Class with Distinction)
Experience
Visiting Undergraduate Research Fellow & RA Intern
Harvard University & Northwestern University
- Architected a high-throughput spatio-temporal data pipeline in Python to ingest, clean, and process ~20M scanned articles (~350GB) from the American Stories dataset.
- Engineered NLP preprocessing workflows and geo-tagging modules to extract structured location and entity metadata from historical U.S. newspaper archives.
- Developed analytical frameworks to quantify and detect reporting biases and sentiment shifts across regional publications over time.
Research Collaborator (Interpretability & AI Safety)
University of Oxford
- Developed the Patchscope-For-Neuron pipeline to causally test neuron reactivation in Gemma-2-2B by injecting residual stream vectors into new contexts via TransformerLens. Repo
- Integrated Sparse Autoencoders (SAEs) to decompose hidden representations into interpretable features, verifying how neurons respond to intrinsic semantic features rather than original textual context.
- Co-authored "Beyond Auto-Interpretability: Mapping a Hierarchical Math Circuit in Gemma-2-2B via Gradient-Based Patchscopes". Report · Repo
Research Intern
University of Hamburg (HCDS Lab), Germany
- Co-first authored research on the illusion of cross-lingual safety transfer in LLMs, developing a Latent Geometric Framework to evaluate hidden-state refusal representations across African languages using the novel LoDNA dataset.
- Co-Organizer, SemEval 2026 Task 9: led a global competition on multilingual online polarization detection (Top 6 on CodaLab, 1,000+ participants). Website
- POLAR: developed scalable scripts for classification and dataset splitting across 22+ languages (110K+ samples); integrated LLM-based pipelines for few-shot and zero-shot learning on GPU clusters.
- AI4Democracy: built an automated monitoring platform to detect and track antidemocratic narratives across 500+ daily web-scraped sources using RoBERTa and Llama 2 / Gemma 2 pipelines.
- Contributed to the AI-MAP project, supported by the Google Award for Inclusion Research.
Research Intern
National Institute of Technology Agartala
- SemEval-2025 Task 11 (Multilingual Multi-label Emotion Detection): achieved top-10 ranks across multiple languages out of 720 participants — Russian 7th, French 8th, Hausa 8th, Hindi 9th. Paper