I enjoy making things. Here are a selection of projects that I have worked on over the years.
From tokenizer to transformer — a hands-on exploration of how large language models learn, implemented and trained entirely from scratch on TinyStories using PyTorch.
A simulation-based study of the Metropolis-Hastings algorithm using Python. This project explores MCMC sampling from t-distributions and normal distributions with both unconditional and conditional proposals. Includes visual analysis using histograms and QQ plots. Developed as part of a graduate-level course in statistical inference. Highly relevant for Bayesian inference, probabilistic modeling, and ML research.
This project explores the use of Gaussian Process Regression (GPR) to model the kinematics of a robotic arm with eight input parameters (joint angles and link lengths) and a nonlinear output (tip distance from origin). Various kernel functions are evaluated, with performance analyzed via uncertainty plots and mean squared error (MSE) metrics.
This project detects solar flare events in NASA’s solar observation video using a statistical framework based on the Generalized Likelihood Ratio (GLR) test. Principal Component Analysis (PCA) is employed to enhance the Gaussianity of the video data, improving the performance and reliability of the GLR-based change-point detection. The approach offers a lightweight, unsupervised change-point detection algorithm.
