Md Kamrujjaman Mobin

Actively participated in global machine learning and data science competitions, applying advanced modeling and data analysis techniques to real-world challenges.

• Awarded Bronze Medal (Top 6.3%) in Image Matching Challenge 2024 by developing a robust multi-model pipeline utilizing COLMAP, LightGlue, SuperGlue, DKM, and LoFTR for high-accuracy 3D reconstruction.
• Secured Bronze Medal (Top 5.7%) in ISIC 2024 - Skin Cancer Detection by implementing sophisticated deep learning models that effectively combined image and tabular data with advanced augmentation techniques.

Engaged in continuous self-directed learning and advanced problem-solving through competitive programming platforms, enhancing algorithmic and data structure proficiency.

• Solved over 1500 algorithmic problems across leading platforms including Codeforces, CodeChef, and LeetCode, demonstrating strong analytical and problem-solving capabilities.

Led the university's Computer Science and Engineering Department basketball team, fostering a culture of teamwork and competitive excellence.

• Captained the team to three consecutive championship victories, demonstrating strong leadership, strategic planning, and motivational skills.

Represented the Computer Science and Engineering Department in various sports, showcasing versatility and commitment.

• Competed in Basketball, Volleyball, Handball, and Football, exemplifying dedication, teamwork, and competitive spirit.

Contributed to the volleyball team's performance, applying strategic thinking and coordination.

• Guided the team to a runner-up position, highlighting effective strategic leadership and seamless team coordination.

Engaged in fine-tuning various LLMs using advanced techniques like QLoRA and n-bit quantization.

Developed a conversational weather application leveraging a custom Model Context Protocol (MCP) server and client, compatible with Gemini LLM for natural language interactions.

Built a lightweight PDF question-answering system utilizing open-source LLMs within a Retrieval-Augmented Generation (RAG) framework.

Developed a full-stack web application for real-time inventory, sales, and order tracking with responsive UI and persistent storage.

Engineered a modular e-commerce platform featuring RESTful APIs to support product management and supplier delivery workflows.

Designed a classroom simulation system inspired by Google Classroom, applying core Java Object-Oriented Programming (OOP) principles.

Designed an IoT solution for real-time flood monitoring using Arduino, ultrasonic sensors, and ESP8266 module for internet-based alerts.

Recognized for multiple first-author publications presented at prestigious international conferences including CinC-2024, ICCIT-2024, Canadian AI-2025, and COLING-2025, contributing significant research to the AI/ML field.

Awarded Bronze Medal for achieving a top 5.7% ranking globally in a medical imaging competition, demonstrating advanced deep learning and data augmentation strategies.

Awarded Bronze Medal for achieving a top 6.3% ranking globally in a computer vision competition, showcasing expertise in 3D reconstruction techniques.

Achieved 6th place in a national challenge focused on road object detection for autonomous vehicles, demonstrating advanced computer vision and machine learning skills.

First Author. Developed an advanced ensemble model integrating RoBERTa, OpenAI detector, and BERT for highly accurate multilingual text detection. Achieved F1-scores of 0.7458 (12th place) for English and 0.7513 (4th place) for multilingual text.

First Author. Developed an ensemble of RoBERTa-base and OpenAI Detector with inverse perplexity weighting for robust cross-domain AI-generated text detection. Achieved TPRs of 0.826 (10th place) for non-adversarial and 0.801 (8th place) for adversarial scenarios.

First Author. Developed CardioFormer, an innovative model incorporating Cross-Channel Patching and Self-Attention mechanisms for ECG analysis. Achieved superior AUROC scores: 96.34% (MIMIC-IV), 89.99% (PTB-XL), and 95.59% (PTB).

First Author. Created CricPose360, a comprehensive dataset, and developed a pose-enhanced YOLOv8 model for real-time cricket shot detection. Achieved 92.21% accuracy and 92.29% F1-score, significantly improving baseline performance by 3%.

First Author. Engineered hybrid UNet-LSTM and ResNet-LSTM models for robust ECG conversion and multilabel classification. Secured 11th place in digitization and 10th place in classification tasks globally.