Project-tutorials | Tanvir Ahmed Sijan

Implementing Transformers from Scratch

Sun, 02 Aug 2026 00:00:00 +0000

This is a work-in-progress project where I build a complete Transformer pipeline from the ground up. The goal is to implement one component at a time, in order: tokenizer → embeddings → positional encoding → attention → feed-forward → layer norm → encoder block → full model → loss → training → evaluation.

Implementing DPO from Scratch

Mon, 22 Jun 2026 00:00:00 +0000

An implementation of Direct Preference Optimization (DPO) for language model alignment using PyTorch. without Hugging Face TRL.

Crash Severity Analysis with Public Geospatial Data

Sat, 22 Feb 2025 00:00:00 +0000

In this project, I demonstrate how publicly available geospatial data and Google Streetview Images can be utilized to determine the correleation between fatal crashes and driver’s visual and built environment for data-scarce regions.

I designed a pipeline that collects Point-of-Interest data from Overture Maps, collects information about road and traffic from Google Traffic Color Codes, weather data from Open-Meteo API and then perform Semantic Segmentation to identify Streetscape Complexity, Street Openness, Building View Index, Greenview Index. Then train four ensemble models (XGBoost, AdaBoost, Random Forest, CatBoost) to predict crash severity. Finally perform SHAP interpretability to map how individual road segments contrbute to fatal road crashes.