I build trustworthy, local-first infrastructure for AI agents.
My work sits at the intersection of AI systems, developer tooling, security, and observability. I care about software that is fast, inspectable, privacy-aware, and honest about what it measured.
- Agent security — runtime permissions, approval workflows, sandboxing, and audit trails
- Agent observability — local-first usage, cost, provenance, and performance tracking
- AI developer tools — coding agents, verification workflows, MCP integrations, and CLIs
- Decision systems — fast structured inference instead of unnecessary token generation
| Project | What it does |
|---|---|
| agentvault | Runtime permission firewall for AI agents with approvals and tamper-evident audit logs |
| burn | Local-first cost and usage ledger for Claude Code, OpenCode, Codex, and Gemini CLI |
| limen | Opt-in local proxy that measures and inspects what coding agents resend |
| flash-coder | Agentic coding CLI with oracle-scored patches and mutation-tested verification |
| myna | Typed decision engine for choice, score, and yes/no decisions in one forward pass |
| ulp | Privacy-first protocol for portable AI coding usage and cost records |
A large part of my open-source work is upstream contribution rather than maintaining isolated forks.
- Laya — major contributions across batching, the TypeScript SDK, CLI workflows, MCP tools, routing, long-document inference, hooks, and benchmark infrastructure.
- freeCodeCamp — curriculum correctness and documentation fixes.
- MDN — web documentation and browser compatibility fixes.
- facebook/time and facebook/astryx — decoder bounds and frame-encoding guards in the NTP/RTCM stack; module path resolution, time parsing, and CDN asset priming in the design system.
- public-apis — data quality, link maintenance, and schema consistency fixes.
- Open-source agent tooling — fixes and improvements across OpenClaw, OpenWorker, Orca, Headroom, Archify, and ProtocolCanary.
- Local-first by default
- No telemetry unless explicitly enabled
- Security boundaries should be testable
- Benchmarks should be reproducible
- Failures should explain themselves
- Fixes belong upstream