Benchmarking LLM code generation
Leading the design of benchmarks that evaluate the reliability of large-language-model code generation. Collaborative research; manuscript in preparation.
2026 — present · python~/dhaka/bangladesh — assistant professor · researcher
~/campus
Every building is a part of my work. Hover to look inside, click to visit.
tip — the buildings are clickable · keyboard works too (tab + enter)
~/about
I'm an Assistant Professor in the Department of Software Engineering at Daffodil International University, where I've taught since 2017, and Lab In-Charge of the DIU Data Science Lab since 2022. Since 2024 I've also been a Research Assistant at the School of Information Technology, Washington University of Science & Technology (USA), working in computer vision, machine learning, LLMs, and NLP within international collaborations.
My research builds and evaluates machine-learning systems for real-world, high-stakes problems. The thread running through it is rigorous evaluation — designing benchmarks and stress-tests that reveal when models genuinely work rather than only appearing to. I've applied deep learning across healthcare, agriculture, and cyber-security, and my current aim is trustworthy and secure AI: models and evaluation methods robust enough for safety-critical deployment.
That focus grew out of nearly a decade of applied work — 25 peer-reviewed publications, many written with my students. I've built deep-learning detectors for road damage and rooftop solar potential, disease classifiers for tea leaves and medicinal plants, intrusion-detection and phishing-defence models in cyber-security, Bangla NLP and sentiment-analysis systems for Bangladeshi platforms, graph-based multi-omics classifiers, and data-mining studies in healthcare and education. The domains vary; the habit is the same — real data, careful measurement, and results that hold up to a second look.
In the classroom I'm an advocate of hands-on, data-driven learning. Theory matters, but I want every student to leave a session having built something.
~/research
25 peer-reviewed publications · 310+ citations. I build machine-learning systems for high-stakes problems — and design the benchmarks that reveal whether they genuinely work.
Leading the design of benchmarks that evaluate the reliability of large-language-model code generation. Collaborative research; manuscript in preparation.
2026 — present · pythonAn execution-grounded benchmark measuring whether an AI-generated security report actually helps a separate fixer repair the vulnerability — judged by whether a real exploit stops firing without breaking the code. Under anonymous review.
2026 · python · manuscript submittedAn empirical study on lowering the cost of stochastic, SWE-bench-style agentic evaluation via non-uniform run allocation — showing that oracle savings of up to ~5× in tokens are not realizable at practical budgets. Manuscript in preparation.
github → agentic-eval-varianceA multidimensional framework for evaluating deep skin-lesion classifiers across calibration, uncertainty, and explanation quality — finding that the most accurate model is often the least trustworthy. Ongoing research.
2026 — present · pytorch · grad-cam++ · mc-dropoutA computer-vision system that segments etioplasts and their sub-organelles, computes geometric metrics, infers parent–child organelle relationships, and generates LLM-based biological summaries. Collaborative research.
2025 — present · django · yolov8 · opencv · scikit-imageAn end-to-end system pairing object detection (YOLOv8/11/12, Faster R-CNN) with LLM-based report generation, converting road-surface imagery into structured defect reports.
2025 — 2026 · ultralytics · pytorchA curated dataset of regional Bangla dialects paired with standard-Bangla and English translations, supporting dialect NLP and machine translation. Built with bangla.ai.
2023 — 2024 · dataset~/experience
School of Information Technology, Washington University of Science & Technology (WUST), USA — computer vision, machine learning, LLMs, and NLP within international collaborations.
Dept. of Software Engineering, Daffodil International University — research, teaching, and supervision of undergraduate and graduate research producing peer-reviewed publications.
Leading data science research projects, mentoring student researchers, and coordinating collaborations in AI and data-driven applications.
Daffodil International University — curriculum development and research leadership.
Daffodil International University — undergraduate software engineering courses; student research mentorship.
~/teaching
Courses taught: Machine Learning, Data Science, Data Structures & Algorithms, Software Engineering, Web Programming, and Software Security — with a build-first philosophy. I'm also building AI Atlas, an interactive platform for learning machine learning by doing.
~/contact
Research collaboration, speaking, student mentorship — my inbox is open.
schedule meetingDepartment of Software Engineering · Daffodil International University · Dhaka, Bangladesh