Staff Engineer @ Google · local-first products · agent systems · agent evaluation · self-improving agent harnesses
Main focuses · Selected builds · How I build
| Local-first products Useful tools that keep personal data and decisions close to the user. |
Agent systems Durable workflows, tools, and control surfaces for coding agents. |
Agent evaluation Evidence, benchmarks, and review loops that make capability visible. |
Self-improving harnesses Agent infrastructure that learns from its own work without hiding the trail. |
Build principle: evidence → local control → human review → shipped artifact.
|
Flying Pig AI Supervised AI customer-service advocate that pauses at structured decision checkpoints before consequential actions. |
mychart-cli Public local-first CLI for exporting patient-owned MyChart records through a Chrome/Puppeteer harness. |
|
Perks Reminder Open-source benefits tracker for welcome bonuses, annual fees, recurring credits, certificates, and loyalty expirations. |
FreshPowder Ski and snowboard deal aggregator tracking current and historical deals across North American retailers. |
|
CastForge Source-to-podcast automation framework with reusable pipeline stages, audio integrations, publishing, and validation. |
Nitan Podcast Weekly Chinese podcast generated from USCardForum discussions and shipped through RSS and podcast platforms. |
|
Context Harness Tiny filesystem memory layer for coding agents: durable context files, companion skills, and local scripts. |
Agent Nexus Manifest-driven deployment of skills, hooks, and MCP servers across modern coding environments. |
|
CPA Phoenix RECENT CLIProxyAPI plugin for igniting fresh Codex accounts and supervising invalid-account recovery through a preserved repair queue. |
AI Builder Brief RECENT Source-linked AI developments turned into evidence-gated review artifacts, with the public feed held closed while editorial policy is tuned. |
|
Continuity LOCAL · IN PROGRESS Workflow, evaluation, repair, checkpoint, and human-gate control plane for a local agent-first research system. |
OpenMuseCall LOCAL · IN PROGRESS Open, supervised calling agent prototype for authorized Mac-to-iPhone cellular calls, with explicit control and safety gates. |
source → build → evaluate → inspect → improve → ship
I run an evaluation loop: turn the goal into a repeatable check, inspect failures and evidence, improve the system, then run the check again before shipping.
Claude Code and Codex are my primary agent tools. I use shell, browser, test, and review tools around them, with Trellis as the primary framework for my agent harnesses.
I care about systems that leave a trail: what happened, what was proposed, what a person approved, and what finally shipped.
More about the portfolio
The through-line is practical automation with visible boundaries: local data when it should stay local, source trails when decisions matter, and human checkpoints before consequential actions.