About
I am a Lecturer in the Department of Software Engineering at Daffodil International University, Dhaka, and a remote Research Assistant at Concordia University supervised by Dr. Yuhong Yan. I received my B.Sc. in Software Engineering from Shahjalal University of Science and Technology in 2025.
My work sits at the intersection of AI for Software Engineering and Trustworthy AI. I study LLM-based software testing and debugging, using program analysis and mutation testing to make AI-assisted testing more targeted and less wasteful, and I evaluate how reliably LLM systems align with expert judgment. I also work on low-resource NLP, building Bangla benchmarks and auditing how well LLMs judge low-resource text.
I am currently applying for PhD positions.
Research Interests
AI for Software Engineering
LLM-based software testing and debugging. I use program analysis and mutation testing to make LLM-generated test suites targeted and cost-aware instead of wasteful.
Trustworthy AI
Auditing LLM systems on the axes that decide whether they can be relied on: expert alignment, evidence grounding, score calibration, ranking reliability, and robustness.
LLM Evaluation for Low-Resource Languages
Benchmark construction and LLM-as-a-judge reliability for Bangla, measuring how closely model judgments track human annotation on text that is thinly represented in pretraining data.
Papers and preprints
R. R. Siam*, S. I. Mugdha*, M. K. Kamrul, A. S. Azad, and S. S. Chowdhury. (*Equal contribution.)
Developed BanglaEduInclusive, a benchmark of 3,877 human-annotated sentences, and evaluated GPT-4o, Gemini-1.5-Flash, and Mistral Large 2.1 for inclusiveness detection using advanced prompting. GPT-4o with chain-of-thought reached 0.763 accuracy and 0.787 F1.
Clus-Test: A Cluster-Aware Multi-Layer Framework for Cost-Efficient Mutation-Guided LLM Unit Test Generation
Author list omitted to preserve double-blind review.
A cluster-aware multi-layer pipeline for Python unit test generation that treats mutation feedback as a compression problem, grouping surviving mutants and escalating only those left unresolved. On 182 EvoEval problems across four semantic-altering subsets it matches the baseline's mutation score (94.3% vs. 93.7%, Wilcoxon p = 0.21) while cutting per-task cost by 67% on average.
Trustworthy LLM-Assisted Venture Screening: Auditing Expert Alignment, Evidence Grounding, and Cross-Domain Calibration
Author list omitted to preserve double-blind review.
An evidence-grounded LLM pipeline for venture screening, audited for expert alignment, evidence grounding, calibration, and robustness. Accepted at the AI in Private Markets workshop at ICAIF '26; being prepared for submission to Canadian AI 2027.
Experience
Teaching
Service
Skills
Get in Touch
Feel free to reach out about research, teaching, or collaboration. Email me at mugdhasamiul@gmail.com and I will respond whenever I can.
Dhaka, Bangladesh
+880 1962-353752