I build ML systems from scratch — then I build products with them.

Co-founder & Backend Developer · letmecookonline.com

A deployed cooking web application. I designed and built the Express.js REST API: authentication with JWT, user profiles, dietary preferences, recipe-generation workflows, and image uploads backed by AWS S3. Also contributed to the Next.js frontend and collaborated through Git/GitHub.

• Express.js
• Node.js
• JWT
• AWS S3
• Next.js
• REST APIs

Autograd engine built from scratch

A scalar automatic-differentiation engine with dynamic computation graphs and reverse-mode backpropagation, implemented from first principles in Python. Covers the chain rule, gradient flow, and a small neural network library trained on top — no PyTorch, no shortcuts.

• Python
• Backpropagation
• Computation graphs
• Neural networks

Regression study with honest evaluation

A small regression study on the classic tips dataset, built around disciplined evaluation rather than model complexity. Every model is scored against a dummy baseline, compared under 5-fold cross-validation with leak-free preprocessing pipelines, and judged by paired fold-wise differences — which show that no richer feature set beats the single-predictor model once fold variance is accounted for.

• Python
• scikit-learn
• pandas
• Cross-validation
• Jupyter

Character-level language models, step by step

A series of character-level language models of increasing complexity: from bigram-count baselines to multilayer perceptrons with learned embeddings, built in Python and PyTorch. Focused on what actually happens inside a language model — loss surfaces, initialization, activation statistics — rather than calling a library.

• Python
• PyTorch
• Language models
• Embeddings
• MLPs