Stony Brook, New York, United States
I am a graduate student at Stony Brook University studying Data Science. Over the past few years, I’ve gained hands-on experience with real-world problems—first as a Data Scientist at Deloitte, where I built data-driven solutions for business challenges like clustering leads, demand forecasting, analyzing customer complaints with NLP, building statistical models to detect early warranty issues, and developing an LLM-powered supply chain assistant. After that, I worked as a data analyst at Stony Brook Medicine, analyzing wearable data to help researchers understand sleep patterns. Recently, at Coforge, I’ve been focusing on building a root cause analysis system for large-scale microservices using graph neural networks to identify failures. I’m also fine-tuning lightweight language models and integrating them with a graph-based retrieval system to suggest incident resolutions. What excites me is working on practical problems where AI can assist organisations in making better decisions or simply automating processes. I’m looking for roles where I can continue building and learning—whether that’s ML engineering, AI research, or any data-driven team where I can make an impact alongside smart, curious people.
Working as a Data Analyst at the Center for Understanding Biology Using Imaging Technology (CUBIT) Lab under Prof. Christine DeLorenzo, analyzing actigraphy monitor data using R programming to study physical activity and sleep patterns.
- Created and processed 50,000 distributed traces by simulating requests across 20+ microservice architectures and injecting 200+ faults using chaos engineering on AWS EKS. Automated data collection and preprocessing with AWS OpenTelemetry, producing a high-quality dataset for developing and evaluating RCA system. - Engineered a GNN-based root cause analysis system for microservice failures. Reduced ML inference by implementing HDBSCAN clustering to group similar anomaly traces. Applied counterfactual analysis to identify root causes through trace restoration simulation and normalized outcome predictions. - Evaluated the RCA system on two public benchmarks and internal datasets, achieving a 0.91 F1 score and 0.88 accuracy, surpassing rule-based baselines in fault localization. - Fine-tuned a Small Language Model with LoRA, cutting fine-tuning compute cost by 68%. Deployed a graph-enhanced RAG system using LightRAG, boosting retrieval accuracy and relevance by 25–40%. - Leveraged vLLM for high-throughput inference and containerized with Docker, achieving 2–5x faster deployment.