Millend Roy

Hello! At Columbia, I am advised by the amazing Prof. Agostino Capponi. My research spans statistics, machine learning, and algorithm design, with a focus on efficient computation and decision-making under uncertainty. I develop mathematical models and scalable algorithms for problems in machine learning and energy systems, and study the physical and economic factors that shape electricity markets.

Previously, I was a SCAI (Societal impact through Cloud and Artificial Intelligence) Research Fellow at Microsoft Research India, where I worked on Project Vasudha. I am fortunate to have been advised by Dr. Akshay Nambi, Tanuja Ganu, and Dr. Shivkumar Kalyanaraman. I applied reinforcement learning and metaheuristic optimization to distributed energy management, worked on EV range anxiety, and helped turn research prototypes into early-stage solutions for Microsoft partners.

I received my Bachelor’s degree in Electrical Engineering from the Indian Institute of Technology (ISM), Dhanbad, where I developed my foundation in power systems. I am especially grateful to Dr. Soham Dutta and Dr. Bhukya Krishna Naik, with whom I worked on my early research projects and whose guidance motivated me to pursue graduate studies.

Outside research, I enjoy playing cricket 🏏, table tennis 🏓, and pool 🎱. I also love sketching ✏️ - you can find some of my work in the gallery. I have formal training in harmonium 🎶 and am now trying to learn the keyboard, with a little more practice this time! :)

selected publications

AlphaSearch: Agentic AI for Price Arbitrage in Energy Systems

Millend Roy and Vladimir Pyltsov

In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, Republic of Korea, 2026

Abs DOI Video Code

Wholesale electricity prices are highly volatile due to fluctuations in weather, renewable generation, demand, and grid conditions. These fluctuations create arbitrage opportunities for energy storage, but realizing this value requires dispatch decisions that account for price uncertainty, battery constraints, and deployment complexity. Existing approaches often require practitioners to separately manage forecasting, optimization, feasibility constraints, and evaluation, making deployment fragmented and hard to interpret. We propose ?search, an agentic AI framework for battery arbitrage that accounts for price uncertainty in wholesale electricity markets. ?search supports and benchmarks heuristic, optimization-based, reinforcement learning, and LLM-assisted decision-making approaches across multiple Independent System Operator (ISO) datasets. We further conduct a series of stress tests to assess the effectiveness of these methods in realistic market settings.
Machine learning based adaptive fault diagnosis considering hosting capacity amendment in active distribution network

Sourav Kumar Sahu, Millend Roy, Soham Dutta, and 2 more authors

Electric Power Systems Research, 2023

Abs DOI

Augmentation of distributed energy resources (DERs) safely in distribution system termed as hosting capacity (HC) is one of the prominent needs to achieve energy sufficiency with minimum emission. However, any amendment in HC over premeditated injection sets up challenges in perspective of situational awareness (SA) of networks for precise decision-making related to fault prediction and location. In this work, authors propose histogram-based gradient boost (HGB) algorithm, an accurate machine learning (ML) technique for fault type detection and location. Due to the unique characteristic of noise cancelation, spectral-kurtosis is utilized for extraction of features of the faulted transient signals. For improved competence of the process, optimized feature importance values are considered. In order to study the efficacy of the proposed method, HC of the network is altered, leading to up-gradation of network parameters. These upgraded parameters are used for retraining the proposed ML algorithm for desired SA, with perception, comprehension, projection, and accurate decision making. The authors also considered other ML techniques to showcase a comparative study with the HGB. The entire analysis is tested on reconfigured IEEE-33 bus distribution system developed in Typhoon HIL real-time simulator. The proposed methodology is also meticulously compared with existing literature to establish its excellence.