resume

Research Assistant - Computer Vision, Multimodal AI & Responsible AI

University of Florida

Advisor: Prof. Yonghui Wu. Designed multimodal retrieval and responsible AI evaluation systems for visual-textual retrieval and grounded generative AI.

• Designed and evaluated multimodal retrieval pipelines combining visual and textual signals; improved NDCG@10 by 21% through dense transformer embeddings, FAISS vector search, and learned reranking.
• Built claim-level verification and context-grounding strategies for generative AI outputs, reducing hallucination rate by 14%.
• Developed responsible AI evaluation harnesses benchmarking faithfulness, accuracy, and safety across model variants.
• Engineered a two-agent LangGraph orchestration system with autonomous task decomposition and tool-use interfaces; reduced average response latency by 32% in a production-ready Ray Serve deployment.

Machine Learning Engineer - Personalization, Ranking & Production ML

ICICI Bank Pvt Limited

Built recommendation, ranking, and experimentation workflows for production banking personalization systems serving 1M+ daily requests.

• Built ML-powered recommendation and ranking systems serving 1M+ daily requests; applied transformer-based ranking on GPU infrastructure, achieving a 9% CTR lift and 12% improvement in recommendation relevance.
• Developed user understanding and segmentation models using supervised and unsupervised learning, sequence modeling, clustering, and anomaly detection.
• Built end-to-end Python training, serving, and experimentation workflows with FastAPI, Docker, and AWS ECS.
• Integrated AI tooling with large-scale data platforms for automated policy personalization from data engineering and training through inference and product.

Research Intern - Sequence Modeling & Behavioral Prediction

Polytechnique Montreal

MITACS Globalink Fellow focused on neural sequence models, temporal pattern learning, and optimized training systems.

• Trained neural sequence models over 10M+ data points for temporal pattern learning; accelerated training 3.8x via parallel computing and optimized C++ pipelines.
• Implemented GPU-accelerated training loops and bottleneck profiling for production-grade behavioral prediction systems.

Research Intern - Neural Retrieval & Multimodal Representation

IIT Patna AI-ML-NLP Lab

Built dense retrieval, reranking, clustering, and GPU inference workflows for large-scale content matching.

• Built dense retrieval and reranking pipelines for content matching over 500K+ documents; improved NDCG@10 by 17% over BM25 using transformer embeddings.
• Developed unsupervised clustering and anomaly detection models over behavioral event streams.
• Optimized GPU-based inference pipelines, reducing end-to-end query latency by 40%.

Master of Science

University of Florida

Computer Science, Machine Learning Track

• Machine Learning
• Deep Learning
• Natural Language Processing
• Large Language Models
• Computer Vision
• Distributed ML
• Statistical Modeling
• Agentic AI Systems

Anisotropic Noise Injection for Improving Utility in Differentially Private SGD

University of Florida

Under review. Contribution: originated the core idea of shaping noise along gradient covariance eigenvectors and led theoretical derivation and empirical evaluation.

Structure-Grounded Medical QA: RDF Retrieval and Claim-Level Verification for Faithful Answering

ACL 2026 Workshop on Structured Understanding and Reasoning for Generative LLMs (SURGeLLM)

Submitted. Contribution: proposed claim-level verification as the core faithfulness mechanism, designed the structured retrieval pipeline, and led evaluation design.

From Sights to Insights: Towards Summarization of Multimodal Clinical Documents

ACL 2024 Long Paper, Main Conference

Contribution: designed the vision cross-attention fusion module and ran ablation studies isolating the contribution of multimodal grounding.

Visual & Multimodal Retrieval System with Responsible-AI Evaluation

Designed a multimodal retrieval pipeline combining vision and text encoders with dense FAISS indexing over 1M+ vectors; achieved 8-12% gains in Recall@100 and 15-20% reduction in p95 latency via GPU-optimized Ray Serve batching.

• Integrated claim-level faithfulness verification and responsible AI evaluation harnesses.
• Benchmarked retrieval quality, latency, and failure modes across model variants.

User Behavior Segmentation & Predictive Profiling Pipeline

Applied supervised and unsupervised learning to model user behavioral patterns and benchmarked XGBoost, transformer, and LSTM architectures across latency and accuracy trade-offs.

• Built temporal feature engineering and drift-detection harnesses.
• Applied clustering, anomaly detection, and sequence classification for user understanding and recommender systems.

ACL 2024 Long Paper

Association for Computational Linguistics

Published "From Sights to Insights: Towards Summarization of Multimodal Clinical Documents" at ACL 2024.

MITACS Globalink Research Fellowship

MITACS Canada

Selected for the MITACS Globalink research program.