Md. Mahmudul Hasan

Research Assistant at CAIL | CSE Graduate, University of Dhaka

I'm a Research Assistant at the Cognitive Agents and Interaction Lab (CAIL). I obtained my B.Sc. in Computer Science and Engineering from the University of Dhaka. My focus of research is foundation models, large language models, vision-language models, and multimodal learning.

Email: mahmudulyeamim [at] gmail [dot] com

Text-to-SQL evaluation remains largely reference-centered. Execution Accuracy (EX), the dominant metric, evaluates a predicted SQL query by comparing its execution result against that of a single reference SQL. However, EX frequently fails when a prediction follows the same intended logic as the reference but differs in underspecified output behavior, when the question or schema admits multiple plausible interpretations, or when the reference SQL itself is noisy or incorrect. We therefore propose Q-SEM, a fine-grained, Question-grounded Semantic Evaluation Metric that evaluates whether a predicted SQL query preserves the intent of the natural-language question. Q-SEM combines two gold-SQL-agnostic views that assess whether the prediction covers and correctly composes the question's semantic requirements with two gold-SQL-contrastive views that determine whether semantic differences constitute actual semantic errors with respect to the question intent. A final adjudicator resolves disagreements among the four views and produces the final acceptability decision. On the expert-labeled ROSE-VEC-BIRD benchmark, Q-SEM achieves 85.61% Cohen's kappa, obtaining the strongest agreement with human judgments and consistently outperforming deterministic metrics (EX reaches only 43.56%) and LLM-based methods across every evaluated LLM.

Despite Bengali being the sixth most spoken language in the world, handwritten text recognition (HTR) systems for Bengali remain severely underdeveloped. The complexity of Bengali script--featuring conjuncts, diacritics, and highly variable handwriting styles--combined with a scarcity of annotated datasets makes this task particularly challenging. We present GraDeT-HTR, a resource-efficient Bengali handwritten text recognition system based on a Grapheme-aware Decoder-only Transformer architecture. To address the unique challenges of Bengali script, we augment the performance of a decoder-only transformer by integrating a grapheme-based tokenizer and demonstrate that it significantly improves recognition accuracy compared to conventional subword tokenizers. Our model is pretrained on large-scale synthetic data and fine-tuned on real human-annotated samples, achieving state-of-the-art performance on multiple benchmark datasets.

January 2026

Semantix

C++ Python FastAPI Next.js MPI OpenMP OpenCV

Built a scalable image segmentation system for generating segmentation masks using K-Means clustering with distributed and shared-memory parallelism.
Integrated MPI-based master-worker scheduling with Bully-algorithm for leader election, and used OpenMP to accelerate clustering on multicore CPUs.

May 2025 - July 2025

Retail Price Forecasting

Python PyTorch Matplotlib NumPy

Built a cleaned, machine-readable price dataset by processing 800+ government bulletins using OCR and manual validation.
Developed Informer-based models for long-horizon forecasting of five essential commodities.

October 2024 - January 2025

Flash: Fastest-Way-to-Learn

Spring Boot Next.js PyTorch

Developed an end-to-end video summarization system leveraging multimodal deep learning techniques.
Built a fully functional course-selling platform.

April 2024

ConvoCorner

Python FastAPI PostgreSQL

Developed real-time messaging with text and image support using WebSockets.
Implemented TCP/IP protocol from scratch to handle reliable data transmission.

September 2023 - November 2023

KodeShell

Android Studio Java

Built an Android app to unify contest schedules and user stats across platforms.
Added community posts, user search, and messaging for community interaction and problem discussion.

February 2023 - May 2023

SmartAcademicManager

Java JavaFX SQLite

Built a desktop app for managing schedules, to-do lists, attendance, and course details.
Designed a user-friendly modern JavaFX interface for efficient academic task management.

October 2022 - December 2022

Cannon Fight

C/C++ SDL2

Developed a 2D projectile-based shooting game in C/C++ using SDL2 with single-player and multiplayer modes.

# Md. Mahmudul Hasan - Research Assistant, Cognitive Agents and Interaction Lab (CAIL), University of Dhaka - CSE Graduate, University of Dhaka ## About I recently completed my B.Sc. in Computer Science and Engineering from the University of Dhaka, where I developed a strong foundation in both theoretical and applied computer science. I am deeply interested in recent advancements in Artificial Intelligence. I currently work as a Research Assistant at the Cognitive Agents and Interaction Lab (CAIL) under the supervision of Dr. Md. Mosaddek Khan. My research interests broadly include Foundation Models, Large Language Models (LLM), Vision-Language Models (VLM), and Multimodal Learning. Beyond research, I have a strong passion for algorithmic problem solving and devoted much of my early undergraduate years to competitive programming. I have also aimed to build a solid understanding of software engineering through academic projects. ## Recent News - June 2026: Our paper **Q-SEM: Fine-Grained Question-Grounded Semantic Evaluation for Text-to-SQL** is currently under review. - February 2026: I joined the Cognitive Agents and Interaction Lab at the University of Dhaka as a Research Assistant. - February 2026: I successfully defended my undergraduate thesis at the University of Dhaka. - September 2025: Our paper GraDeT-HTR was accepted at EMNLP 2025 System Demonstrations. Paper: https://arxiv.org/abs/2509.18081. Code: https://github.com/mahmudulyeamim/GraDeT-HTR. - June 2025: We released GraDeT-HTR for public use. Project site: https://cognistorm.ai/hcr. ## Publications ### Q-SEM: Fine-Grained Question-Grounded Semantic Evaluation for Text-to-SQL - Authors: **Md. Mahmudul Hasan**, Meherun Farzana, Mehrajul Abadin Miraj, Aniket Joarder, Mahmudul Hasan, Abir Chakraborty Partha, Md. Ahasanul Alam, Md. Tanvir Alam, Redwan Ahmed Rizvee, Md. Fahim Arefin, Md. Mahmudur Rahman, Md. Mosaddek Khan - Status: Under Review (EMNLP 2026) - Contribution: Reframed Text-to-SQL evaluation around intent preservation, overcoming limitations of execution-based correctness metrics. - Contribution: Built multi-view LLM adjudication, achieving 85.61% Cohen's kappa on ROSE-VEC-BIRD. - Abstract: Text-to-SQL evaluation remains largely reference-centered. Execution Accuracy (EX), the dominant metric, evaluates a predicted SQL query by comparing its execution result against that of a single reference SQL. However, EX frequently fails when a prediction follows the same intended logic as the reference but differs in underspecified output behavior, when the question or schema admits multiple plausible interpretations, or when the reference SQL itself is noisy or incorrect. We therefore propose Q-SEM, a fine-grained, Question-grounded Semantic Evaluation Metric that evaluates whether a predicted SQL query preserves the intent of the natural-language question. Q-SEM combines two gold-SQL-agnostic views that assess whether the prediction covers and correctly composes the question's semantic requirements with two gold-SQL-contrastive views that determine whether semantic differences constitute actual semantic errors with respect to the question intent. A final adjudicator resolves disagreements among the four views and produces the final acceptability decision. On the expert-labeled ROSE-VEC-BIRD benchmark, Q-SEM achieves 85.61% Cohen's kappa, obtaining the strongest agreement with human judgments and consistently outperforming deterministic metrics (EX reaches only 43.56%) and LLM-based methods across every evaluated LLM. ### GraDeT-HTR: A Resource-Efficient Bengali Handwritten Text Recognition System utilizing Grapheme-based Tokenizer and Decoder-only Transformer - Authors: **Md. Mahmudul Hasan**, Ahmed Nesar Tahsin Choudhury, Mahmudul Hasan, Md. Mosaddek Khan - Venue: EMNLP 2025 Demo - Paper: https://arxiv.org/abs/2509.18081 - Code: https://github.com/mahmudulyeamim/GraDeT-HTR - Contribution: Utilizes a decoder-only architecture for Bengali handwritten text recognition without any pretrained language model. - Contribution: Employs a grapheme-based tokenizer, which significantly improves recognition accuracy for Bengali script. - Abstract: Despite Bengali being the sixth most spoken language in the world, handwritten text recognition (HTR) systems for Bengali remain severely underdeveloped. The complexity of Bengali script--featuring conjuncts, diacritics, and highly variable handwriting styles--combined with a scarcity of annotated datasets makes this task particularly challenging. We present GraDeT-HTR, a resource-efficient Bengali handwritten text recognition system based on a Grapheme-aware Decoder-only Transformer architecture. To address the unique challenges of Bengali script, we augment the performance of a decoder-only transformer by integrating a grapheme-based tokenizer and demonstrate that it significantly improves recognition accuracy compared to conventional subword tokenizers. Our model is pretrained on large-scale synthetic data and fine-tuned on real human-annotated samples, achieving state-of-the-art performance on multiple benchmark datasets. ## Education ### B.Sc. in Computer Science and Engineering - University of Dhaka - Duration: 2022 - 2026 - CGPA: 3.95/4.00 - Supervisor: Dr. Md. Mosaddek Khan - Location: Dhaka, Bangladesh ## Work Experience ### Research Assistant - Cognitive Agents and Interaction Lab (CAIL), University of Dhaka - Duration: February 2026 - Present - Location: Dhaka, Bangladesh - Supervisor: Dr. Md. Mosaddek Khan - Building Text-to-SQL systems and relational foundation models for business intelligence applications. ### Undergraduate Research Assistant - Cognitive Agents and Interaction Lab (CAIL), University of Dhaka - Duration: January 2024 - January 2026 - Location: Dhaka, Bangladesh - Supervisor: Dr. Md. Mosaddek Khan - Worked on efficient Bangla handwritten text recognition and reasoning in vision-language models. ### Physics Instructor - Udvash Academic and Admission Care - Duration: January 2022 - December 2023 - Location: Dhaka, Bangladesh - Taught physics to grades 9-12 students and university admission candidates. ## Selected Projects ### Semantix - Date: January 2026 - Technologies: C++, Python, FastAPI, Next.js, MPI, OpenMP, OpenCV - Link: https://github.com/mahmudulyeamim/Semantix - Built a scalable image segmentation system for generating segmentation masks using K-Means clustering with distributed and shared-memory parallelism. - Integrated MPI-based master-worker scheduling with Bully-algorithm for leader election, and used OpenMP to accelerate clustering on multicore CPUs. ### Retail Price Forecasting - Date: May 2025 - July 2025 - Technologies: Python, PyTorch, Matplotlib, NumPy - Built a cleaned, machine-readable price dataset by processing 800+ government bulletins using OCR and manual validation. - Developed Informer-based models for long-horizon forecasting of five essential commodities. ### Flash: Fastest-Way-to-Learn - Date: October 2024 - January 2025 - Technologies: Spring Boot, Next.js, PyTorch - Link: https://github.com/saged-sama/Flash---Fastest-Way-to-Learn - Developed an end-to-end video summarization system leveraging multimodal deep learning techniques. - Built a fully functional course-selling platform. ### ConvoCorner - Date: April 2024 - Technologies: Python, FastAPI, PostgreSQL - Link: https://github.com/mahmudulyeamim/ConvoCorner - Developed real-time messaging with text and image support using WebSockets. - Implemented TCP/IP protocol from scratch to handle reliable data transmission. ### KodeShell - Date: September 2023 - November 2023 - Technologies: Android Studio, Java - Link: https://github.com/mahmudulyeamim/KodeShell - Built an Android app to unify contest schedules and user stats across platforms. - Added community posts, user search, and messaging for community interaction and problem discussion. ### SmartAcademicManager - Date: February 2023 - May 2023 - Technologies: Java, JavaFX, SQLite - Link: https://github.com/mahmudulyeamim/SmartAcademicManager - Built a desktop app for managing schedules, to-do lists, attendance, and course details. - Designed a user-friendly modern JavaFX interface for efficient academic task management. ### Cannon Fight - Date: October 2022 - December 2022 - Technologies: C/C++, SDL2 - Link: https://github.com/mahmudulyeamim/Cannon-Fight - Developed a 2D projectile-based shooting game in C/C++ using SDL2 with single-player and multiplayer modes. ## Competitive Programming - March 2023: ICPC Asia Dhaka Regional Onsite Contest 2022, 51st out of 158 teams, youngest student representing my university in 2023. - February 2023: ICPC Asia Dhaka Regional Preliminary Online Contest 2022, 32nd out of 1,477 teams. Qualified for the Regional Onsite Contest; selected to represent my university in my sophomore year as the only student from my cohort. - February 2023: CodeChef Max Rating: 1962, Username: mahmudulyeamim. Profile: https://www.codechef.com/users/mahmudulyeamim - January 2023: Codeforces Max Rating: 1582, Username: mahmudulyeamim. Profile: https://codeforces.com/profile/mahmudulyeamim ## Contacts - Email: mahmudulyeamim [at] gmail [dot] com - GitHub: https://github.com/mahmudulyeamim - LinkedIn: https://www.linkedin.com/in/mahmudulyeamim/ - X: https://x.com/mahmudulhyeamim - Google Scholar: https://scholar.google.com/citations?user=F7JM9l0AAAAJ

Recent News

Publications

Q-SEM: Fine-Grained Question-Grounded Semantic Evaluation for Text-to-SQL

Md. Mahmudul Hasan, Meherun Farzana, Mehrajul Abadin Miraj, Aniket Joarder, Mahmudul Hasan, Abir Chakraborty Partha, Md. Ahasanul Alam, Md. Tanvir Alam, Redwan Ahmed Rizvee, Md. Fahim Arefin, Md. Mahmudur Rahman, Md. Mosaddek Khan

Reframed Text-to-SQL evaluation around intent preservation, overcoming limitations of execution-based correctness metrics.

Built multi-view LLM adjudication, achieving 85.61% Cohen’s κ on ROSE-VEC-BIRD.

GraDeT-HTR: A Resource-Efficient Bengali Handwritten Text Recognition System utilizing Grapheme-based Tokenizer and Decoder-only Transformer

Md. Mahmudul Hasan, Ahmed Nesar Tahsin Choudhury, Mahmudul Hasan, Md. Mosaddek Khan

Utilizes a decoder-only architecture for Bengali handwritten text recognition without any pretrained language model.

Employs a grapheme-based tokenizer, which significantly improves recognition accuracy for Bengali script.

Education

B.Sc. in Computer Science and Engineering, University of Dhaka

CGPA: 3.95/4.00

Supervisor: Dr. Md. Mosaddek Khan

Work Experience

Research Assistant

Cognitive Agents and Interaction Lab (CAIL), University of Dhaka

Building Text-to-SQL systems and relational foundation models for business intelligence applications.

Supervisor: Dr. Md. Mosaddek Khan

Undergraduate Research Assistant

Cognitive Agents and Interaction Lab (CAIL), University of Dhaka

Worked on efficient Bangla handwritten text recognition and reasoning in vision-language models.

Supervisor: Dr. Md. Mosaddek Khan

Physics Instructor

Udvash Academic and Admission Care

Taught physics to grades 9–12 students and university admission candidates.

Competitive Programming

ICPC Asia Dhaka Regional Onsite Contest 2022

51st out of 158 teams, represented my university in my sophomore year as the only student from my cohort.

ICPC Asia Dhaka Regional Preliminary Online Contest 2022

32nd out of 1,477 teams, qualified for the Regional Onsite Contest

CodeChef Max Rating

1962, Username: mahmudulyeamim

Codeforces Max Rating

1582, Username: mahmudulyeamim