Projects
Most of these started as a gap in my own understanding. Each one is a short case study: why I built it, what I actually did, and what it taught me — with a link to the code.
A from-scratch implementation and experimentation sandbox for denoising diffusion models in PyTorch — forward noising, the reverse denoiser, and the sampling loop, derived by hand.
A research project asking whether a structural causal model in the latent space of a bidirectional GAN can disentangle protected attributes well enough to generate minority-group images that fix a biased training set.
A single reference implementation of GANs, VAEs, and normalizing flows built side by side to compare how each family trades off sample quality, likelihoods, and training stability.
A from-scratch Vision Transformer for image classification, built to understand patch embeddings, attention over image patches, and exactly how much data attention needs to beat a convolutional baseline.
An implementation of the Transformer architecture from first principles, with experimentation on sequence tasks, built to internalise attention rather than recite it.
A network-science study of academic collaboration built from real DBLP bibliographic data: graph construction, centrality, and community detection used to study how a department's research reputation grew over time.