Prasanjit Dubey

I am a Ph.D. candidate in Machine Learning at Georgia Tech's H. Milton Stewart School of Industrial & Systems Engineering, where I am fortunate to be advised by Prof. Xiaoming Huo. My research builds statistical methods with provable guarantees for two intertwined settings: multiple hypothesis testing with fast, scalable algorithms for exact error control, and federated, communication-constrained learning (knowledge distillation and conformal coverage for large language models and LLM swarms under tight bandwidth budgets). A growing part of my work sits at their intersection: error control that stays valid when evidence is distributed across sites or accumulates sequentially over time.

During my Ph.D., I worked as a quantitative researcher at Bloomberg and a research scientist at AT&T Labs. Before that, I was a senior quantitative analyst building credit-risk models at ANZ and a research intern at Walmart Labs. I earned my master's in Applied Statistics and Informatics at IIT Bombay, where I graduated first in my class.

I'm always glad to talk about multiple testing, federated learning, or a possible collaboration. Feel free to email me at dubey.prasanjit@gmail.com.

Research Interests

• Multiple Hypothesis Testing
• Conformal Prediction
• Anytime-Valid Inference
• Federated Learning
• Knowledge Distillation
• Large Language Models
• LLM Swarms
• Retrieval-Augmented Generation
• Statistical Learning Theory
• Family-Wise Error Rate
• Communication-Constrained Learning
• Optimization

More on the four research lines →

Selected Papers

All publications →

1. BOOST: Power-Optimal Strong-FWER Testing for Block-Structured Multiplicity

   Prasanjit Dubey and Xiaoming Huo

   Accepted NeurIPS · 2026 Paper

2. Matching Rates and Optimal Allocation for Federated Probe-Logit Distillation under Heterogeneous Bandwidth Budgets

   Prasanjit Dubey and Xiaoming Huo

   Under review Invited submission, Journal of the American Statistical Association (Theory & Methods) · 2026 Paper

   ★ Poster and Travel Award, Conference on Statistical Foundations of AI (CSFAI), University of California, Santa Barbara, 2026

   ★ Poster presented at STAI-X, Harvard University, 2026

3. Most Powerful Tests with Exact Family-Wise Error Rate Control: Necessary Conditions and Fast Computation for General K

   Prasanjit Dubey and Xiaoming Huo

   Revision under review Journal of Computational and Graphical Statistics (JCGS) · 2025 Paper

   ★ Invited talk, Optimisation for Statistics and Learning session, INFORMS Annual Meeting, 2025

4. Federated Multiple Testing with Quantized p-Values under Family-Wise Error Control

   Prasanjit Dubey and Xiaoming Huo

   Under review Bernoulli · 2026 Paper

   ★ Accepted at the NeurIPS 2026 Second Workshop on ML×OR: Mathematical Foundations and Operational Integration of Machine Learning for Uncertainty-Aware Decision-Making

5. LLmFPCA-detect: LLM-powered Multivariate Functional PCA for Anomaly Detection in Sparse Longitudinal Texts

   Prasanjit Dubey, Aritra Guha, Zhengyi Zhou, Qiong Wu, Xiaoming Huo, and Paromita Dubey

   Preprint arXiv · 2025 Paper

   ★ 2nd Place Paper Award, Student Research Showcase, INFORMS Regional Analytics Conference (IRAC), 2025

Research & Professional Experience

Alongside my dissertation research, I have worked as a quantitative researcher in industry and on sponsored research at Georgia Tech, building statistical and machine-learning systems that shipped to real users.

Industry

• Summer 2025

  Quant PhD Intern, Quantitative Risk Analytics

  Bloomberg L.P. · New York, USA

  Built a Kalman-filter hazard-curve model for pricing credit instruments from latent Nelson–Siegel factors and macroeconomic covariates, integrated into Bloomberg Terminal's Credit VaR engine; implemented a credit-rating-migration module using macro-conditioned multinomial logit and intensity-based models to simulate obligor- and index-level rating paths.

• Summer 2024

  Research Scientist Intern, Data Science & AI Research

  AT&T Labs, Inc. · New Jersey, USA

  Developed a framework integrating large language models with multivariate functional data analysis for sparse longitudinal customer-journey text, now an arXiv preprint; shipped a production anomaly-detection application (Emotional AI) for behavioral interventions.

• 2019 – 2021

  Senior Quantitative Analyst, Wholesale Credit Risk Modeling

  Australia and New Zealand Banking Group (ANZ) · Bengaluru, India

  Built the Financial Ratio Surveillance Tool (FRST), a production R Shiny stress-testing application forecasting default probabilities and assessing the predictive power of financial ratios; led the bank's statistical COVID-19 impact assessment across commercial portfolios for APRA regulatory requirements.

• Summer 2018

  Research Intern, Data Science (International Pricing & Demand Forecasting)

  Walmart Labs · Bengaluru, India

  Built a two-stage price-elasticity model to forecast category-level sales from item-, region-, and covariate-specific price elasticities: k-means clustering, then an ensemble of ridge, lasso, and elastic-net models fit within each cluster, reducing mean absolute percentage error by 42.4% relative to the incumbent method.

Sponsored Research

• Jan – Sep 2026

  Graduate Research Assistant, Georgia DOT Sponsored Research

  Georgia Institute of Technology · Atlanta, GA, USA

  Developed a statistical framework establishing how many valid days a traffic-monitoring station needs in a month with incomplete data before its monthly average can be published, and applied it across Georgia's network of 233 continuous count stations. Delivered to transportation planners as the Data-Sufficiency Explorer, an interactive dashboard. PI: Prof. Xiaoming Huo.

  Code

Education

• Aug 2021 – May 2027 (expected)

  Ph.D. in Machine Learning

  Georgia Institute of Technology · Atlanta, GA, USA

  • H. Milton Stewart School of Industrial and Systems Engineering; minor in Mathematics
  • Advisor: Prof. Xiaoming Huo
  • GPA 3.85 / 4.00
• 2017 – 2019

  M.Sc. in Applied Statistics & Informatics

  Indian Institute of Technology Bombay · Mumbai, India

  • Department of Mathematics · Class Rank 1 of 38 · GPA 8.96 / 10.00
  • Thesis: Financial Data Mining Using Multivariate Copulas
• 2014 – 2017

  B.Sc. in Statistics, First Class Honors

  University of Calcutta · Kolkata, India

  • Department of Statistics; minor in Mathematics & Computer Science
  • Grade 92.5%

Technical Skills

Programming

Python, R, R Shiny, C++, SQL, MySQL, MATLAB

Machine Learning

LLMs, PyTorch, scikit-learn

Computing & Tools

High-performance computing (Slurm), GPU computing, Apache Hadoop (MapReduce, Streaming), Git, LaTeX

Graduate Coursework

Mathematical Foundations of Machine Learning, High-Dimensional Probability, High-Dimensional Statistics, Statistical Inference, Time Series Analysis, Stochastic Processes, Convex Optimization, Stochastic Optimization, Computational Data Analysis