website · optimizer research · code intelligence · agent coordination
I build AI-native systems: research prototypes, developer tools, and product-grade web software. My current work is centered on optimizer efficiency, local-first code intelligence, and coordination protocols for AI coding agents.
I care about the whole path from idea to evidence: implement the sharp version, measure it honestly, and turn the parts that survive into usable systems.
| Thread | What I am testing |
|---|---|
| Optimizer research | Whether Muon-style optimizers can preserve sample-efficiency gains with fewer Newton-Schulz iterations. |
| Agent tooling | Local codebase context, symbol-aware coordination, and conflict detection before agents touch the same code. |
| AI-native products | Interfaces where models are part of the system architecture, not a feature added at the end. |
| Project | Notes |
|---|---|
| fastmuon-local-research | Local PyTorch research lab for Muon-family optimizers. First signal: FastMuon-NS3/NS4 preserved much of Muon's small-LLM gain with fewer Newton-Schulz iterations. |
| graft | Local-first codebase context engine for AI coding tools via MCP. TypeScript, static analysis, tree-sitter, dependency graphs. |
| wit | Agent coordination protocol: declare intents, lock symbols, and detect conflicts before code is written. |
| awesome-cli-coding-agents | Curated directory of terminal-native AI coding agents and orchestration harnesses. |
| harmoniq-site | Product/frontend work in TypeScript. |
| Area | Tools |
|---|---|
| AI / research | Python, PyTorch, TensorFlow, NumPy, scikit-learn, NLP |
| Systems | TypeScript, Rust, C++, Java, tree-sitter, MCP, CLIs |
| Product | Next.js, Svelte, Vue, Supabase, PostgreSQL, HTML/CSS |
Prototype quickly.
Prefer measurements over aesthetics.
Keep the interface simple enough to survive real use.The best starting point is amaarmc.org. For code, research, and open-source work, the relevant projects are linked above.