Software engineer with 12 years of experience in web development (Rails / React / AWS), currently expanding into machine learning and deep learning.
An empirical study of neural scaling laws applied to SVG code generation. Trained GPT-style models (1.3M-88M params) on 107M tokens of SVG data and fit power-law curves to characterize how loss scales with model size. Compared Standard Parameterization with muP for zero-shot learning rate transfer.
Tech: Python, PyTorch, muP, BPE tokenization, power-law fitting