Resume Resume

Ph.D.

Carnegie Mellon University

Information Systems

M.S.

Carnegie Mellon University

Information Systems

B.S.

University of California, San Diego

Applied Mathematics | Business

Applied Scientist Intern

Amazon

Led end-to-end development of a generative recommendation system by pre-training a custom user encoder on 172k interaction sequences and progressively aligning it with LLMs, achieving strong gains over prompting baselines (Hit@3 60.1%, NDCG@3 67.5%) with 93.4% fewer tokens and 64.8% lower cost. Built a production-ready RAG/search pipeline with FAISS HNSW over 60k user embeddings and implemented distributed multi-GPU training with FSDP and BF16 mixed precision to scale training and inference efficiently.

Data Scientist

HireBeat

Leveraged machine learning and NLP techniques to develop a data analysis pipeline that optimized content recommendation algorithms, significantly improving personalization accuracy and user engagement.

Founder

R-Eat

Entrepreneurship: Founded and grew an online food delivery business by enhancing customer experience through UI optimization and personalized marketing, resulting in a 60% sales increase and a loyal customer base of over 2,000.

TCS Presidential Fellowship

Carnegie Mellon University

CMLH Fellowship in Generative AI in Healthcare

Center for Machine Learning and Health, Carnegie Mellon University

CMU Heinz College Research Excellence Award

Center for Machine Learning and Health, Carnegie Mellon University

Agentic Q&A Systems for Trustworthy Organ Transplantation Guidance

• Q&A System: Developed Q&A systems for organ transplantation patient care by integrating retrieval-augmented generation (RAG) with adaptive frameworks that synthesize patient-specific medical history, clinical guidelines, and dynamic risk factors to deliver tailored guidance and support.
• RAG/Retriever: Engineered a hybrid retrieval pipeline combining both sparse and dense retrievers (DPR, ANCE) to accurately extract and verify medical content from over 100 transplant-specific handbooks.
• LLM Reasoning Framework: Developed an incentivized reasoning mechanism that optimizes LLM performance by rewarding multi-step chain-of-question sequences, enabling the system to reason freely and proactively request critical diagnostic information.

LLM Alignment & Multi-Turn Consistency

• Distributed Training: Architected and optimized distributed training infrastructure for LLMs (Mistral, LLaMA, Deepseek) across 4 nodes with 16 NVIDIA GH200 GPUs, implementing tensor parallelism, pipeline parallelism, and ZeRO optimizer techniques to efficiently scale model training and reduce computational overhead.
• Fine-tuning: Performed fine-tuning of state-of-the-art LLMs both locally (on-premises clusters) and via OpenAI's official GPT-4o API, successfully improving model consistency, confidence-aware reasoning, and achieving state-of-the-art performance.
• Evaluation: Proposed a metric to quantitatively assess multi-turn response stability, prioritizing early interaction accuracy and swift recovery from initial errors. Released a comprehensive benchmark to test the model's ability to sustain coherent and contextually accurate responses during dynamic, multi-turn conversations.

Chronic Kidney Disease Progression Analysis with DL & XAI

• Deep Learning: Developed Temporal-Feature Cross Attention Mechanism (TFCAM), an attention-based deep learning framework that explicitly captures temporal and feature-level interactions to enhance clinical predictive modeling for Chronic Kidney Disease (CKD) progression.
• XAI: Provided multi-level explainability including temporal insights, feature importance ranking, and cross-temporal feature interactions, enabling clinicians to interpret model predictions transparently and identify actionable clinical insights.
• Enhanced Representation & Delivery: Designed an interactive, user-friendly dashboard GUI accessible to diverse user groups; integrated LLMs to automatically generate clear, insightful summaries of predictions and model interpretations, enhancing usability and clinical decision-making support.