AI

I started with ChatGPT and GitHub Copilot in early 2024, using them for coursework, writing and debugging code, and side projects. Most of my building now happens in Claude Code, which I have used since 2026.

I did not want to only use these tools, so I learned how they work. I watched 3Blue1Brown, Andrej Karpathy's lectures and Hugging Face's course videos.

I can explain transformers and attention, tokenization, embeddings, and how agents decide on and call tools.

I have worked hands-on with chunking and embeddings. My RAG Database Research project is a tested learning playbook plus a planned local RAG design for a healthcare referral system. That full pipeline is designed, not built.

In my phone agent, the Python code (not the model) searches a database for troubleshooting steps.

I plan before I build: a design, then a plan, then tests and code. I split work across sub-agents, keep reusable skills, hooks and memory, and run tests and automated review before anything ships.

I also keep a graphify map of each codebase and query it first, so the agent reads only what matters.

Superpowers: the brainstorm, plan, test-first and verify-before-done workflow.

UI/UX Pro Max: a design database of styles, palettes and UX guidelines that I use across my web projects, including this site.

Rederive: a skill by Eric Zou that teaches a codebase by asking questions until you re-derive its design decisions yourself. I use it to test whether I understand the systems I build with AI, not just ship them.

Thorough code review: a skill I wrote. It runs one agent per area of a project, verifies every finding, and produces a severity-ranked PDF with a fix for each.

I keep a COG vault and an Obsidian vault. At the end of a session I save a braindump of decisions, lessons and to-dos. Each project has its own log and notes, and I keep class and learning notes there too.

Claude reads and writes the vault, so context carries from one session to the next.