Rakib Ullah

Status: Submitted to Computer Methods and Programs in Biomedicine

Key Contributions:

• Led pioneering research on clinical outcome prediction for low-resource medical text settings without native language corpora
• Fine-tuned transformer-based models achieving 65.3% ROC-AUC for length-of-stay prediction and 64.9% ROC-AUC for mortality prediction on 33,000+ clinical records
• Developed end-to-end NLP pipeline for dataset preprocessing
• Co-authored manuscript with publication-ready figures, statistical analyses, and performance tables

Status: Submitted to IC3IT 2026

Key Contributions:

• Conceptualized and led the first research on low-resource Bengali and Bengali-English code-mixed memes
• Developed novel dataset of 3,247 instances with detailed annotation guidelines achieving 79% inter-annotator agreement
• Designed custom fusion model achieving 77.65% Macro F1-score for classifying inflammatory, hateful, and benign memes
• Co-authored manuscript with publication-ready visualizations and detailed result analyses

Focus: Benchmarking PEFT methods and model compression for Bengali NLP

Key Contributions:

• Benchmarking parameter-efficient architectures (LoRA, QLoRA) for Bengali NLP tasks
• Analyzing performance–efficiency trade-offs for real-time production deployment
• Conducting statistical analysis and visualization of model compression techniques
• Authoring research manuscript including methodology, evaluation, and result analysis

Overview: Production-ready RAG system enabling natural language querying of long scientific PDFs

Key Features:

• Built scientific QA chatbot with LangChain leveraging Gemini Pro's long context window
• Enhanced system to interpret figures and tables for accurate, context-aware answers from 20+ page research papers
• Developed microservices architecture (FastAPI + Streamlit) with containerization using Docker
• Deployed on Render with comprehensive testing for reliability and maintainability

Impact: Addresses Gemini's limited usability for scientific QA in native app

Ranking: 37^th / 151 teams

Score: 81% Macro F1 (Private Leaderboard)

Approach & Contributions:

• Instruction-tuned all layers of Qwen3-VL 8B using LoRA for multimodal classification
• Applied 4-bit quantization via Unsloth to enable training under strict compute constraints
• Handled the complete ML pipeline independently, including preprocessing, training, validation, and evaluation
• Addressed significant distribution shift between training and test data to improve generalization

Affiliation: Sylhet Engineering College