# S M Nafim Niloy

> S M Nafim Niloy is a CSE undergraduate and ML researcher in Dhaka working on graph neural networks, medical imaging, computer vision and Bengali NLP.

Computer Science and Engineering undergraduate with research experience in graph neural networks, computer vision, medical imaging, and natural language processing, seeking opportunities to apply machine learning and software engineering skills in a professional setting. Experienced in end-to-end ML pipelines, full-stack development, and reproducible experimentation.

- Current role: Student Research Associate at Smile Care (Oct 2026 – Present)
- Location: Dhaka, Bangladesh
- Email: admin (at) nafim (dot) dev (written this way to avoid spam bots)
- Contact page: https://nafim.dev/#contact
- Website: https://nafim.dev/
- Status: Open to research & job opportunities

## Pages
- [Portfolio](https://nafim.dev/): experience, education, research, papers, projects, skills and contact
- [Blog](https://nafim.dev/blog/): writing on ML research, engineering and life
- [Full profile for AI assistants](https://nafim.dev/llms-full.txt): everything on this site as plain text

## Experience
- Student Research Associate, Smile Care (Oct 2026 – Present)

## Education
- Bachelor of Science in Computer Science and Engineering, American International University-Bangladesh
- Higher Secondary Certificate, Dhaka College

## Research
- Undergraduate Thesis: Do Heterogeneous Graph Neural Networks Improve Cross-Cohort Glioma Grading? A TCGA-CGGA Benchmark with Graph Ablation and Explainability — Under review at NetMAHIB

## Papers
- Commuting Pattern and Challenges Among Students in Dhaka City: An Analysis for Sustainable Urban Mobility — Conference · ICMAE 2025 · Presented

## Work in progress (not yet published)
- Do Heterogeneous Graph Neural Networks Improve Cross-Cohort Glioma Grading? A TCGA-CGGA Benchmark with Graph Ablation and Explainability — Journal · NetMAHIB · Under Review
- Heterogeneous Lesion Graph Network for Diabetic Retinopathy Grading: An Exploratory Study of Spatial Lesion Interaction Modeling — Journal · JSCDM · Submitted
- Explainable Machine Learning of Texture Signatures in Pancreatic Cancer via Leakage-Safe Radiomics and Rule Mining — Conference · ICEFronT · Accepted
- TeaBD-HybridNet: Maximizing Pretrained Knowledge Transfer in a Dual-Branch CNN-ViT Architecture for Tea Leaf Disease Classification — Conference · ICRPSET 2026 · Submitted

## Projects
- [BanglaLLM](https://demobanglallm.streamlit.app/): A demo version of a 7B model. Developed and trained a 100M-parameter decoder-only language model for Bangla using a custom 16K-vocabulary tokenizer, with training and inference implemented in Python and Transformers.
- [Animal Image Classification with ResNet50](https://animalimageclassificationresnet50-gyvtphv5nivq8wbtedsgcu.streamlit.app/): Built and fine-tuned a ResNet50 model to classify images of 10 animal species, reaching 96.59% validation accuracy through transfer learning, data augmentation, and hyper-parameter tuning. A Streamlit app lets you upload an image and get a prediction instantly.
- [EduTutor BD](https://github.com/smnafimniloy/edu-tutorBD): End-to-end RAG-based AI tutoring system for SSC students using OCR-processed Bengali textbooks, LangChain, and FastAPI, with retrieval-grounded answers and student progress tracking.
- [Stock Prediction Portal](https://github.com/smnafimniloy/stock-prediction-portal): Full-stack stock prediction platform built with Django, React, and Bootstrap, integrating machine learning models for historical market-data analysis, trend prediction, and interactive visualization.
- [Smart Cafe](https://github.com/smnafimniloy/SmartCafe): Web application built from scratch in PHP and MySQL following an MVC architecture, with role-based dashboards for Admin, Manager, and Customer and full CRUD for menus, inventory, and customer preferences.

## Blog posts
- [Building BanglaLLM: Training a Language Model from Scratch](https://nafim.dev/blog/building-banglallm-from-scratch/): Why I trained a 100M parameter transformer specifically for Bangla, what broke along the way, and what I learned about tokenisation, data quality, and the gap between "it runs" and "it works".
- [A Weekend at Kaptai Lake](https://nafim.dev/blog/weekend-at-kaptai-lake/): Leaving Dhaka behind for a few days, I found myself on a wooden boat gliding across Kaptai Lake — the largest artificial lake in Bangladesh — surrounded by mist, green hills, and an unexpected kind of quiet.

## Elsewhere
- [GitHub](https://github.com/smnafimniloy)
- [LinkedIn](https://linkedin.com/in/nafimniloy)