I build fast, understandable software across AI systems, developer tools, accessibility, music tech, and product experiments. My favorite projects have a simple test: the internals are interesting, and the result is useful enough that someone would actually run it.
Right now I am especially interested in:
- Small language models and transformer training loops from scratch.
- Product-quality AI interfaces that are open, hackable, and cheap to run.
- Accessibility tooling, maps, and civic software with real-world constraints.
- Systems projects where performance, correctness, and readability all matter.
| Project | What it is | Stack |
|---|---|---|
| army | A compact GPT-style transformer training stack written from scratch. | C++17, Accelerate, CPU kernels |
| zllm | Building an LLM in C++ to understand the whole stack. | C++ |
| free3chat | Open source AI chat interface inspired by fast modern chat products. | TypeScript |
| spotipi | Home lab music streaming service built for a campus cybersecurity club. | Svelte |
| minesweep | Route-safety app for high-risk travel contexts. | TypeScript |
| mappi | Public accessibility map database for safer, easier travel. | TypeScript |
| Area | Current shape |
|---|---|
| AI systems | Training loops, tokenizers, inference experiments, and C++ model internals. |
| Product work | Chat interfaces, maps, music tools, and small apps with real users in mind. |
| Systems work | CPU kernels, gradient checks, build tooling, and readable low-level code. |
| Open source | Public repos as working notebooks: demos first, polish where it matters. |
- Start small, keep the core readable, and add polish where users feel it.
- Prefer working demos over vague architecture.
- Write code that can be debugged at 2 a.m.
- Keep performance work grounded in measurement, not vibes.