🎓  M.S. Computer Science · Columbia University _{(2026 – 2027)}
🏛️  B.Tech CSE · VIT Vellore _{(9.15 CGPA · ranked 6th in program)}
🛠️  Building LLM agents, RAG pipelines and evaluation harnesses
📦  Shipped aiollm-gateway on PyPI
📄  Published at IEEE ICoECIT 2026
💬  Ask me about LangGraph, LLM evals, synthetic data

AI Research Intern · Jio Institute, Navi Mumbai · Jun 2026 to Jul 2026

• Built automated LLM experiment pipelines with LangChain, LangGraph and the OpenRouter API, running controlled evaluations on GSM8K and Visual7W
• Engineered a multi-turn Solver–Prober setup to test whether LLMs stay consistent under adversarial questioning, across different prompting tones
• Extended multiple-choice evaluation experiments to vision-language models by building visual reasoning benchmarks

Undergraduate Research Assistant · VIT Vellore · Advisor: Prof. Gopichand G · Jul 2025 to May 2026

• Built LLM-guided data augmentation to reduce bias in AI systems, with checks on semantic fidelity, safety and statistical validity
• Designed experiments measuring fairness and subgroup performance gaps; contributed to an IEEE-published paper

LGSA: LLM-Guided Synthetic Augmentation for Mitigating Bias in AI Systems
IEEE ICoECIT 2026 · arXiv:2510.13202
_{Uses an LLM to generate label-consistent counterfactual samples for underrepresented groups, validated in multiple stages. Narrows subgroup performance gaps without hurting overall accuracy.}

📚 ragpipe Production RAG with enforced citations. Answers questions over 40 arXiv papers, and every claim must trace to a retrieved passage, or the system refuses. Hybrid BM25 + dense retrieval, cross-encoder reranking, a LangGraph pipeline, and a RAGAS quality gate on every pull request.	⚡ llm-gateway Async rate-limit layer for LLM APIs, published on PyPI as aiollm-gateway. Paces requests under provider limits, retries what's worth retrying, batches prompts, and fails over with a circuit breaker. Zero runtime dependencies.
🧪 Synthetic Data Quality Validator Scores synthetic tabular data on fidelity, utility, privacy, diversity and calibration, detecting numeric and categorical features automatically. FastAPI backend in Docker, Streamlit front end.	🔍 AI Hallucination Taxonomy Explorer Detects and categorizes LLM hallucinations with heuristic and LLM-based pipelines. Measures prompt–output drift with sentence embeddings, and its taxonomy refines itself as new cases come in.

Languages
LLM & AI
Data
Engineering

_{Also: retrieval-augmented generation, prompt engineering, LLM evaluation, synthetic data.}