Md Siam Ahmed

I am an instructor for General Physics 2 (Phys 1125)

I am the TA for Foundation of Computer Science 2

I work on Applied Machine Learning and Scientific Computing.

The Smart File Organizer is an automated desktop application that intelligently organizes cluttered folders (particularly Downloads folders) by analyzing both file metadata and content. Unlike traditional file organizers that rely solely on file extensions, this system uses content analysis to understand what files actually contain and categorizes them accordingly. What problem does it solve? Most computer users struggle with messy download folders and file systems that become increasingly disorganized over time. After 6 months to a year, manually organizing hundreds or thousands of files becomes an overwhelming task. Users often have files with non-descriptive names (like "document.pdf" or "IMG_1234.jpg") that provide no context about their actual content. Example Problem: A file named "abcd.pdf" could be an invoice from September, a recipe, or a work document. Traditional organizers would only see it as a PDF file, but our system will read its content, recognize it as an invoice from September, and automatically place it in an "Invoices/2025/September" folder structure.

This study investigates how dataset design and parameter-efficient fine-tuning (PEFT) methods influence the performance of large language models (LLMs) in specialized non-English domains. Using a comparative literature-based methodology, the research evaluates two distinct adaptation strategies: LoRA-based fine-tuning on DeepSeek R1 Distill 8B and LLaMA 3.1 8B using a 64.6k-entry Turkish professional Q&A dataset, and multi-task fine-tuning of the Qwen-7B model using a heterogeneous Chinese electric-power corpus derived from academic papers, patents, and enterprise documents. Results from the Turkish study show substantial accuracy gains, particularly for DeepSeek, with improvements of up to 10% in weaker domains. The Chinese study demonstrates broader multi-task enhancements, including a 9% increase in classification accuracy, a 10.5-point gain in NER F1 score, a 58.6% rise in BLEU-4 for question–answering, and significant improvements in summarization metrics. Comparative analysis indicates that LoRA enables efficient, domain-targeted adaptation, while structured, multi-source datasets yield deeper semantic generalization across diverse tasks. Together, these findings show that optimal domain adaptation requires aligning dataset structure with fine-tuning strategy, emphasizing that high-quality dataset engineering is as critical as the choice of PEFT method. The study highlights practical pathways for developing specialized, cost-efficient LLMs for non-English professional and scientific applications.