I'm an AI Researcher based in Pune, India, working at the intersection of Technical AI Governance, adversarial testing, and safety evaluation of frontier models. My research focuses on developing systematic approaches to ensure AI systems are safe, aligned, and beneficial.
I hold a Master's in Computer Applications and a BSc in Physics from Savitribai Phule Pune University. I was selected among 350+ global applicants for the ICTP Advanced ML Workshop in Italy, and have completed AI Safety training through BlueDot Impact.
20th International Conference on Computing and Machine Learning (CML) · March 2024
Computing and Machine Learning, Volume 2 - Springer Nature
US export controls on AI fail because they exempt publicly-released models. Anyone can download and fine-tune open-source models for military/harmful use at 1/1000th the original training cost. The report argues governments should focus on what's actually controllable: the expensive computer chips for training, transparency requirements, legal liability for developers, and defensive investments.
Building production-ready AI pipelines, developing goal-based AI Agents using A2A protocol and Google ADK, implementing RAG systems with various vector databases.
Led research on prompt engineering, fine-tuning, and RAG architectures. Conducted comparative analysis of LLMs (GPT-4, Claude-3.5, Gemini Pro, Llama-3).
Built POC for RAG architecture using Pinecone. Integrated Hugging Face libraries and researched optimization techniques for LLMs.
Selected among 350+ global applicants (1 of 2 from India). Engaged with Intel, Nvidia, IBM experts on GNNs, Transformers, Neuromorphic Computing.
Savitribai Phule Pune University · 2022-2024
GPA: 8.27/10 (3.31/4)
Savitribai Phule Pune University · 2018-2022
GPA: 8.46/10 (3.39/4)
New Delhi, India · December 2025
BlueDot Impact · Dec 2024
BlueDot Impact · Dec 2024
ADaSci · Aug 2024
March 2024
IUCAA · Dec 2023
LIGO-IUCAA · May 2023
Control Evaluations Framework
AI controls evaluation demonstrating how different Claude models respond to deceptive audit instructions. Main task: identify vulnerability and classify it. Side Task: find a back door and misclassify it. Tested 3 frontier models across 2 tasks to measure strategic misclassification behavior.
GitHub →Frontier Models Security Assessment
Systematic adversarial testing of GPT-5, Claude Sonnet 4.5, and Gemini 2.5 Pro. Revealed 69pp security gap with capability exploitation (51% success) as primary vulnerability.
GitHub →Biomedical QA (OpenAI Paper)
Reproduced OpenAI's methodology using GPT-2 and PubMedQA. Few-shot achieved 32.5% higher performance and 57.1% improvement on cross-domain transfer.
GitHub →Llama-2 Fine-tuning
Implemented Quantized LoRA for efficient fine-tuning of Llama-2-7b-chat-hf with Open-Platypus dataset, reducing computational resources significantly.
GitHub →ML-based Route Planning
Optimized marine routes using Copernicus Marine data. Achieved 5% reduction in fuel costs and CO₂ emissions using Linear Regression and Random Forest.
GitHub →Read my research and analysis on AI governance, safety, and policy:
Deep dives into AI policy, governance frameworks, and technical safety research. Exploring the intersection of AI capabilities and regulatory mechanisms.
Visit Substack →Technical Blog
Technical Blog
Technical Blog
Technical Blog
Medium
Why Claude or ChatGPT can read "Cun yew ndirstawnd mei cenTense?" as "Can you understand my sentence?" Initially it seems magic as every letter is scrambled or replaced, yet Claude or basically any LLM instantly reconstructs the intended meaning. This ability reveals how large language models actually work.
Medium
A lot of tech is currently being used in making hiring easy and seamless, one such aspect is Voice Analysis which plays a very important role. Voice analysis contains 48 different dimensions of vocal emotions categorization and there much more to it.
Medium
Guided and evaluated semester projects for 10+ First and Second year Masters students, providing mentorship in machine learning implementation, research methodology, and project development. Oversaw end-to-end project execution from ideation to final presentation, ensuring academic rigor and practical applicability.
Demonstrated comprehensive methodologies for leveraging AI tools to enhance research workflows and quality. Covered advanced techniques for literature review automation, data analysis acceleration, and research paper writing assistance. Emphasized ethical considerations and best practices for AI-augmented academic research.
Mentored engineering interns in building production-grade Generative AI applications, designing and implementing ML pipelines, and developing autonomous AI Agents. Provided hands-on training in modern AI frameworks, RAG architectures, and agentic systems. Guided technical problem-solving and professional development in AI engineering practices.
Delivered comprehensive introduction to LLM architectures, training methodologies, and practical applications. Covered transformer models, attention mechanisms, prompt engineering, and fine-tuning strategies. Included hands-on demonstrations of working with state-of-the-art language models and discussed real-world deployment considerations.
Explained cutting-edge developments in AI Agent architectures, autonomous decision-making systems, and multi-agent coordination. Covered agent-based planning, tool use, ReAct frameworks, and the latest research in agentic AI systems. Demonstrated practical implementations and discussed integration strategies for enterprise applications.
Mentored graduate physics students in conducting variable star observations and photometric analysis for their Master's thesis projects. Provided guidance on observational astronomy techniques, data collection protocols at the observatory, CCD imaging, light curve analysis, and scientific writing. Supervised thesis development from observation planning through final defense.
Amateur Astronomer: Executive Member at India's oldest astronomy organization with 4+ years of active involvement