Research Activity

My research focuses on offline & hierarchical reinforcement learning for Intelligent Tutoring Systems, robust evaluation methods (DR/WDR), and sequence modeling. I work on building end-to-end ML systems and deploying them in real-world educational settings.

Current Research

Offline Reinforcement Learning for Intelligent Tutoring Systems

I combine Conservative Q-Learning (CQL) with policy distillation from reference "good" policies to improve tutoring effectiveness. Using Doubly Robust (DR) and Weighted Doubly Robust (WDR) Off-Policy Evaluation (OPE) methods on a 10-year dataset with 12 problems and a problem→step hierarchy, I achieved an improvement in outcome prediction AUC by +0.06. This work addresses the challenge of learning effective tutoring policies from historical student interaction data without requiring online exploration.

Meta-Reinforcement Learning Transfer (PEARL)

I adapt underperforming tutor policies using limited new interactions from multiple prior tutors, reducing cold-start regret in offline simulations. This work enables knowledge transfer across different tutoring contexts, allowing new tutors to benefit from the experience of previously trained systems.

Hierarchical Reinforcement Learning & Delayed Rewards

I trained and deployed a CQL-based high-level tutor that operates in a hierarchical framework to handle delayed rewards in educational settings. The deployed system showed statistically significant improvement in natural learning gain, demonstrating the effectiveness of hierarchical approaches for long-term educational outcomes.

Student Modeling with Transformers and Time-Series

I build transformer and time-series models to predict post-test scores and learning gain from student trajectory logs. These models help understand student learning patterns and enable personalized educational interventions.

Previous Research

"A System for Providing Transparency in Government Transaction using Private Blockchain"

We used blockchain to store important government documents for safekeeping. In this paper we addressed the problem of corruption in Bangladesh and how it has a deep effect in public sector of government. We proposed a system using blockchain that will store the transaction of government project and provide transparency and security to the data. Any discrepancy found would be submitted to proper authority. We used private blockchain for our system. Users of our system enter transaction information into our system and later verified from relevant information. Any inconsistency is reported to moderator of the system. We used Hyperledger Fabric and Hyperledger Sawtooth to build our system.