I build open-source tools for AI agents, with a focus on the parts that get hard once several agents work together:
- 🧠 Memory. What should agents share, what must stay private, and when does a fact stop being true?
- 🎮 Agents that build things. Describe a 2D game in one sentence and get a playable Godot 4 game, made by a local fine-tuned model.
- 📏 Evaluation. Benchmarks that measure what agents actually do, run in CI, with honest numbers.
- 🪪 Honest models. LLMs that know their own limits and look things up instead of guessing.
- 💻 Local-first. Everything should work offline on a single GPU, then scale up to cloud models.
I work in public: every release ships with its benchmark results, including the numbers that don't flatter me yet. Contributions and real-world failure cases are always welcome.
