Carnegie Mellon University

M.S. in Machine Learning, Machine Learning Department (SCS) — Pittsburgh, PA

• GPA: 4.0
• Coursework: Advanced Machine Learning, Deep Learning Systems, Multimodal Machine Learning, Diffusion and Flow Matching

Indraprastha Institute of Information Technology (IIIT) Delhi

B.Tech. in Computer Science and Artificial Intelligence — New Delhi, India

• CGPA: 9.50/10.00 — Department Rank 2, Institute Rank 3
• Dean's List Award for Academic Excellence in all semesters
• Teaching Assistant for Machine Learning (postgraduate)
• Coursework: Data Structures, Algorithm Design and Analysis, Database Management Systems, Object-Oriented Programming, Operating Systems, Data Science, Convex Optimization, Machine Learning, Natural Language Processing, Information Retrieval, Deep Learning, Large Language Models, Reinforcement Learning

Amazon Web Services

Applied Scientist Intern, AWS Agentic AI — Seattle, WA

• Improved instruction following in multi-agent systems by developing context routing strategies for efficient information selection and agent coordination, increasing task reliability across complex enterprise workflows.
• Enhanced agent reliability by integrating verification through autoformalization and LLM-based judges, and trained domain-specific policy models using GRPO to optimize agent behaviour for analytics tasks.

Microsoft Research India

Research Engineering Intern, Extreme Classification Group — Bangalore, India

Supervisors: Dr. Amit Sharma, Dr. Manik Varma

• Built HORIZON, a large-scale user modeling benchmark from long-range cross-domain histories, with optimized big-data handling (Apache Parquet, SQLite). Proposed a FAISS-accelerated dual-encoder retrieval setup to evaluate LLM-based recommenders.
• Devised a multi-step chain-of-thought framework with GPT-4o for Bing Ad recommendations, improving interestingness by 15% and relevance by 5%. Benchmarked SLMs and reasoning models on HORIZON, scaling inference and training on distributed multi-GPU setups (vLLM, Axolotl).
• Led the migration of Bing Ads recommendations from GPT-4o to lightweight 8B models, designing a synthetic reasoning chain generation and task-specific distillation pipeline, achieving a 100× inference speedup at teacher-level recall and precision.

IBM Research India

Research Engineering Intern — New Delhi, India

Supervisors: Dr. Sameep Mehta, Dr. Nishtha Madaan

• Developed lightweight LLM guardrails for enterprise applications: blacklisting user-suggested topics (87% accuracy via LDA and keyword extraction), toxicity detection (91–95% with a bi-encoder), and RAG-based hallucination detection using Pinecone.
• Designed and deployed a React/JS web GUI for enterprise-safe LLM interactions with REST-based APIs for the guardrails, integrating red-teaming and evaluation workflows for factual and domain grounding. Accepted for demo at AAAI 2024.

Carnegie Mellon University

Graduate Research Assistant — Pittsburgh, PA

Supervisors: Dr. Daniel Fried, Dr. Sean Welleck

• Investigating reinforcement learning for efficient reasoning in LLMs and agentic systems, in collaboration with Amazon AGI.

FORUM Lab, Carnegie Mellon University

Graduate Research Assistant — Pittsburgh, PA

Supervisor: Dr. Aditi Raghunathan

• Investigated emergent misalignment in instruction-tuned LLMs and the data distributions that aggravate it.

INK Lab, University of Southern California

IUSSTF-Viterbi ML Research Intern — Los Angeles, CA

Supervisor: Dr. Xiang Ren

• Designed MEMOED, the first framework attributing LLM cultural knowledge to memorization versus generalization using distribution analysis, signal-noise metrics and IR techniques, revealing how pre-training data distributions shape memorization and enabling corpus quality and bias assessment. Accepted at ICLR 2025.
• Led the first large-scale analysis of LLM pre-training corpora by scaling MEMOED to billions of documents via sliding-window time optimizations, custom batching for API calls, and parallelized generation-evaluation pipelines, achieving a 50× speedup.

Mila – Quebec AI Institute & McGill University

MITACS Globalink Research Intern — Montréal, Canada

Supervisor: Dr. Jackie Chi Kit Cheung

• Worked on erasure in natural language generation systems, in collaboration with the FATE team at Microsoft Research Montréal.
• Developed a qualitative taxonomy for categorizing erasure, contributing to improved evaluation of harm and reliability in NLP systems.

Infosys Center for Artificial Intelligence, IIIT-Delhi

Undergraduate Researcher — New Delhi, India

Supervisor: Dr. Anubha Gupta

• Designed a prosody-preserving speech-to-speech translation system using Whisper-Fast, Meta's MMS-TTS and X-vector voice cloning, achieving 15% higher MOS and +8 BLEU over direct S2ST systems. Published at ICLR 2024's Tiny Track.
• Proposed a co-attention multi-task learning framework in PyTorch for speaker emotion recognition, improving F1 by 8% on unseen multilingual speaker profiles over state-of-the-art models. Published at INTERSPEECH 2024.
• Built efficient speech-to-speech translation pipelines using Whisper with Flash Attention, and developed a gender bias-reduction method (patent pending) via data and alignment-based interventions.
• Curated a morphology-aware evaluation benchmark to assess translation of grammatically gendered languages in LLMs.

MIDAS Lab, IIIT-Delhi

Undergraduate Researcher — New Delhi, India

Supervisor: Dr. Rajiv Ratn Shah

• Enhanced scientific QA and reasoning in open-source LLMs via retrieval-augmented fine-tuning with curated corpora, improving answer accuracy and BERTScore by 16 points. Published at BDA 2023.
• Proposed an image-text alignment strategy using image descriptions to improve multimodal representations in open-source VLMs, boosting accuracy and reasoning by 7% on benchmarks like MMLU and SciQA.
• Led development of the first Indian textbook-based physics knowledge corpus using OCR, and of text- and image-based Hindi and English scientific question answering datasets.

Amazon ML Challenge 2024 — Multimodal LLM OCR

• Secured a top-10 finish (6th in India) out of 20,000+ teams, earning a presentation opportunity with Amazon scientists. Built a few-shot learning pipeline using ensembled open-source VLMs for entity extraction from 1.2M e-commerce images, achieving a 0.72 F1 score.

Cross-Lingual Transfer Learning for Indian Languages

• Investigated cross-lingual transfer for Indian languages, designing a cross-attention module in PyTorch to boost transfer in low-resource settings. Achieved 10–12% improvement across three downstream cross-lingual tasks.

Multimodal Browser Agent

• Built a JavaScript and CSS Chrome extension with real-time web scraping, Pinecone-based multi-index retrieval and multimodal LLM-based RAG for context-aware Q&A. Enabled personalization via controlled browser-history exposure and multi-tab support through caching with LlamaIndex.

• Google DeepMind Research Symposium 2025 — selected among the top 150 undergraduates in India.
• Amazon ML Challenge 2024 — ranked 6th of 20,000+ teams nationwide, securing an Applied Science internship.
• IUSSTF-Viterbi India Program 2024 — selected among 15 scholars nationwide for a fully funded research internship at USC's INK Lab.
• MITACS Globalink Research Scholarship 2024 — fully funded research internship at Mila-Quebec AI Institute and McGill University.
• Dean's List for Distinguished Academic Excellence 2024 — awarded for top rank (A+) in three courses in one academic year; the only awardee from my batch.
• Amazon ML Summer School 2023 — selected for Amazon's annual summer school.
• Summer Undergraduate Research Fellowship 2023 — awarded by IRD, IIIT-Delhi for research on scientific question answering in LLMs.
• Joint Entrance Examination — JEE Main 99.87 percentile, JEE Advanced 99.76 percentile out of 1.1 million candidates.

• ACM Multimedia 2024, ICLR 2024 (Tiny Track), COLING 2025, ACL 2025, EMNLP 2025.