Focused on Software Engineering, LLMs, and Security. All key contributions are validated with state-of-the-art comparisons.
Secret Breach Detection in Source Code with Large Language Models
October 2024 - July 2025 | Undergraduate Thesis, ESEM 2025 Technical Track Publication
- Introduced a novel approach for Secret Breach Detection using a Small Language Model (SLM) fine-tuned with QLORA.
- Model demonstrably outperforms several established state-of-the-art regex-based tools (like Trufflehog) and large, general-purpose LLMs (like GPT-40).
- Supervisor: Dr. Rifat Shahriyar, Professor, CSE, BUET.
Technology & Tools: QLORA, DeepSeek-7B, Gemma-7B, LLAMA-3.1-8B, Mistral-7B, DeepSeek-V3, GPT-40.
Secret Leak Detection in Software Issue Reports using LLMs: A Comprehensive Evaluation
July 2024 - October 2025 | Undergraduate Thesis, Under Review in MSR 2026, ArXiv
- Presented the first large-scale study and a robust hybrid detection pipeline for secret leaks in GitHub issue reports.
- Curated and released the first public benchmark dataset of over 54,000 labeled instances.
- Demonstrated fine-tuned LLMs achieve state-of-the-art performance (up to 0.945 F1).
- Supervisors: Dr. Rifat Shahriyar (BUET); Dr. Gias Uddin (York University).
Technology & Tools: Small Language Models (RoBERTa, BERT, CodeBERT), QLORA, PEFT, GPT-40, Gemini-2.0-Flash.
BanglaForge: LLM Collaboration with Self-Refinement for Bangla Code Generation
August 2025 - September 2025 | Independent Research Group, Workshop co-located with IJCNLP-AACL 2025
- Introduced a novel framework for generating executable code from Bangla descriptions.
- Utilizes a retrieval-augmented dual-model collaboration paradigm with iterative self-refinement based on execution feedback.
- Achieved a competitive Pass@1 accuracy of 84.00% on the BLP-2025 benchmark.
Technology & Tools: Dual-LLM architecture, Retrieval-Augmented few-shot prompting, Gemini-2.5-Pro.