research.profile("Syed Mehedi Hasan Nirob")
Prospective PhD student and AI engineer. I build production LLM systems and study where AI breaks: detecting AI-generated code and images, evaluating robustness beyond benchmark accuracy, and understanding how AI tools change the work of software engineers.
01 / Publications
Benchmarked handcrafted spatial, texture, and frequency features with classical and gradient-boosted classifiers on CIFAKE, evaluating both detection performance and probability calibration.
Results
Feature fusion lifts F1 by about four points over baseline features.
Compared handcrafted feature models, CodeBERT-based classifiers, and score-level fusion, using feature ablations and dataset checks to probe how reliable detection really is.
Results
Controlled Black formatting and AST normalization tests reveal substantial sensitivity — benchmark accuracy alone overstates reliability.
Developed an SVM classifier for Bangla factoid questions, including collecting the Wikipedia data behind it, and evaluated unigram and question-structure features with five-fold cross-validation.
Results
Best result used the 380 most frequent unigram features.
A survey-based empirical study of how AI coding tools affect engineers across experience levels and organizations. If you write software in Bangladesh, your answers help.
Participate in the survey03 / Research interests
ML/AI systems, software engineering, NLP, generative AI, trustworthy AI, and human-centered systems — with one thread through all of them: can we trust what the model produced?
Detecting AI-generated code and images, and testing whether detectors hold up under realistic shifts.
Evaluation that goes past benchmark accuracy: calibration, robustness, and failure analysis.
Low-resource language work, from Bangla question classification to language identification.
How AI coding tools change productivity, skills, and the quality of engineering decisions.
Interfaces and workflows that help people understand, assess, and trust AI output.
Retrieval, agents, and LLM pipelines that stay reliable in production, not just in a notebook.
04 / Research in practice
Production AI work gives me real settings for studying reliability, retrieval, and how people actually use AI output.
AI Engineer
LLM-based customer service over WhatsApp with product recommendations and structured business-data retrieval, plus privacy-sensitive AI systems across client projects.
Founding Engineer
Built and scaled a production LLM/RAG chatbot platform, working on hallucination reduction, retrieval reliability, response quality, and user trust — and learning human-AI interaction from demos and onboarding.
Pre-RAG data pipeline
Mapped inconsistent source data into a standardized schema and indexed harmonized records in Pinecone with embeddings for downstream retrieval.
Text classification
Multinomial Naive Bayes and Logistic Regression on character and word n-gram features from the DSL dataset.
Edge AI
Collaborated on integrating AI-enabled edge devices, focusing on real-time communication.
Intern
Supported retinal image processing workflows for diabetic screening software — an early look at applied AI in healthcare.
05 / Methods & toolkit
06 / Education & community
Shahjalal University of Science and Technology · 2012 – 2016
Trained university students in competitive programming, after solving 1,000+ problems on UVA, Codeforces, and Topcoder.
Designed problems for 5 national inter-university contests; helped sponsor and judged a national hackathon.
A top-20 finish at the ICPC Asia Dhaka Regional, plus three top-10 finishes in national contests.
A monthly stipend for sustained academic excellence and a first-year university scholarship.