I'm a Computer Science and Engineering student at Chittagong University of Engineering & Technology (CUET), Bangladesh, interested in machine learning, natural language processing, and multimodal AI.
My work includes NLP research, shared-task competitions, deep learning, and building AI-powered applications.
- Natural Language Processing (NLP)
- Machine Learning & Deep Learning
- Large Language Models (LLMs) and RAG
- Multimodal AI and Vision-Language Models
- AI for Healthcare
Enhancing Fine-Grained Arabic Readability Assessment through Ordinal-Aware Transformer Ensembling and Decision Calibration
- Authors: Sumaiya Zaman, Miftahul Jannat Rishta
- Method: Ensemble of MARBERT, CAMeLBERT MSA, AraELECTRA, and AraBERTv2 using ordinal-aware loss functions and validation-based decision calibration.
- Achievement: 3rd place in the Strict Track | Blind Test QWK: 0.848
- Links: Paper · ACL Anthology· Code
Leakage-Controlled MARBERTv2–SVM Ensemble for Arabic Harmful-Prompt Classification
- Authors: Miftahul Jannat Rishta, Sumaiya Zaman
- Method: MARBERTv2 and TF-IDF character n-gram SVM combined through a Logistic Regression stacking ensemble, supported by leakage-controlled evaluation and five rounds of error-driven data augmentation.
- Achievement: 5th place in Subtask B1 | Blind Test Macro F1: 0.860
- Links: Paper · ACL Anthology · Code
Single-Choice Reformulation and Confidence-Calibrated Refinement for Culturally Grounded Hallucination Detection
- Authors: Sumaiya Zaman, Miftahul Jannat Rishta
- Method: Qwen2.5-VL-7B-Instruct with 4-bit quantization, single-choice reformulation, and confidence-gated refinement using seven paraphrased prompts and majority voting.
- Achievement: 6th place in the English Track | Blind Test Accuracy: 0.942
- Links: Paper · ACL Anthology · Code
Gated Cross-Modal Attention with Domain-Adapted Text Encoding for Vaccine-Critical Meme Detection
- Authors: Sumaiya Zaman, Miftahul Jannat Rishta, Shiti Chowdhury
- Method: Twitter-RoBERTa + CLIP ViT-L/14 with gated cross-modal attention.
- Achievement: 3rd place in the shared task | Macro F1: 0.8357
- Links: Paper · DOI · Code
Multi Architecture Transformer Ensemble for Detecting Abusive Tamil Text Targeting Women
- Authors: Miftahul Jannat Rishta, Sumaiya Zaman, Shiti Chowdhury, Hasan Murad
- Method: Heterogeneous ensemble of 12 MuRIL, IndicBERT v2, and XLM-RoBERTa-Large classifiers using soft-voting.
- Achievement: 4th place in the shared task | Macro F1: 0.8086
- Links: Paper · DOI · Code
Multilingual Transformer Based Hope Speech Detection for Coarse and Fine-Grained Classification in Tulu
- Authors: Sumaiya Zaman, Miftahul Jannat Rishta, Shiti Chowdhury, Hasan Murad
- Method: XLM-RoBERTa, MuRIL, and mBERT with soft-voting for coarse-grained classification; XLM-RoBERTa for fine-grained classification.
- Achievement: 1st place in both shared-task subtasks | Macro F1: 0.58 (coarse-grained), 0.42 (fine-grained).
- Links: Paper · DOI · Code
- Competition: Datathon 2.0 — Bengali LLM Hallucination Detection Challenge
- Team: The Villagers
- Approach: Ensemble learning and stacking combining BanglaBERT, mDeBERTa-v3, TF-IDF, and Qwen2.5-7B-Instruct.
- Tech stack: Python, PyTorch, scikit-learn, Hugging Face Transformers.
- Code: GitHub Repository
Contributed to an open-source Retrieval-Augmented Generation (RAG) engine during my industrial attachment at BdREN.
- Contribution: Developed a web crawler and HTML-to-clean-text extraction pipeline, collecting clean documents from BdREN websites.
- Focus: Web crawling, data extraction, and document preprocessing for RAG.
- Organization: BdREN Innovation
- Code: GitHub Repository
A weather application for checking weather conditions.
A project focused on Bangla-language tooltip functionality.
A web application for managing blog content and organizing posts.
A car rental web application for managing rental-related operations.
A machine learning project focused on detecting potentially fraudulent credit card transactions.
A task management application for organizing daily activities and keeping track of tasks.
A personal finance application for tracking budgets and managing expenses.
A programming practice platform designed to support coding and problem-solving in an offline environment.
A campus feedback platform for collecting anonymous feedback and analyzing sentiment to identify trends.
- Languages: Python, JavaScript, TypeScript, PHP, SQL
- AI/ML: PyTorch, TensorFlow, Scikit-learn
- NLP: Transformers, LLMs, RAG
- Web Development: React, FastAPI, Spring Boot
- Tools: Git, GitHub, Kaggle