← Home
Projects
A curated catalog of public code, documented experiments, and selected private systems work. Public projects link to inspectable evidence. Private work is described only at the level needed to explain the problem, result, and what I learned.
Lethe · agent behavioral drift benchmark
An open benchmark for measuring how agent behavior changes during extended operation. Outcomes are checked externally instead of relying only on the agent's own report.
Blackreach · browser and research agent
A local-first agent for stateful browser work, focused observations, resumable sessions, and recovery on real websites. Backed by public code and more than 3,000 tests.
Huginn · self-hosted web collection API
A deployable API for JavaScript-rendered pages and structured extraction, designed around explicit job state, retries, and local operation.
claude-voice · local voice for AI coding
Local text-to-speech with synchronized word highlighting. No cloud API required, packaged for simple installation, and used by other developers.
Rigr · agent regression testing
A focused toolkit for freezing expected agent behavior and detecting change in future runs.
AMP Discovery · protein language model experiment
A leakage-aware antimicrobial peptide classification experiment using ESM-2 and LoRA, with documented evaluation and candidate screening.
ghboard · GitHub dashboard for the terminal
A Go TUI for contribution activity, starred repositories, and notifications, built to keep a common GitHub workflow in the terminal.
Fenrir · Canadian PC-part price tracker
A local price-history tool with confidence scoring, a JSON interface for agents, and a terminal interface for people.
Mimir · a deliberate product reset
A private Linux terminal project. I froze a broad first generation and restarted with a smaller terminal-first boundary. The case study is about honest scope control, not archived features presented as current work.
Private systems work
Recent work includes Rust/Linux applications, multi-agent development infrastructure, and governed local-first workspaces. Sanitized walkthroughs are available without private architecture, data, or operational details.