Customer ops leader who builds AI tools that people can trust. Founder of loafecho.
- print-compliance: screens STL and 3MF files for geometry that resembles regulated weapon parts and returns PASS, REVIEW or BLOCK with a reason a person can check.
- ygo-duel-agent: trains a small language model to play Yu-Gi-Oh on one consumer GPU, using a Rust duel runner on the open-source EDOPro engine and evaluations fixed before each training round.
- surface-pro7-linux-camera: kernel and userspace patches for the Surface Pro 7 cameras and touchscreen driver on Linux, with the fault-injection scripts and offline checkers used to test them.
- glia-agent: a maintenance agent for a Markdown notes vault that can only propose changes; a person approves each one and a separate executor applies it as one revertible commit.
- agent-sandbox: least-privilege microVM sandboxes for AI agents, plus a Rust tool server that holds risky actions until a person approves them.
- homelab-as-code: infrastructure as code for a one-node Proxmox homelab (OpenTofu, Ansible, Argo CD), with firewall rules tested against the ports that are actually listening and a drift gate that answers PASSED, DRIFT or ERROR.
- Checks that must be able to fail. A test only counts if removing the thing it protects turns it red. glia-agent switches off each of its guards in turn and requires the suite to fail; the camera repo's evidence checker recomputes each verdict from the raw test lines instead of trusting the summary.
- Approval before action. In glia-agent and agent-sandbox the agent proposes or requests, a person approves, and only then does anything change. The rules for what an agent may never touch live in code, not in a prompt.
- Honest limits. Every README has a Limits section that says what was not tested and where the tool falls short, such as geometry that can send a file to review but cannot reject it, or patches tested on one device.
