Engineer in Los Angeles / Long Beach. I build software that has to be trusted before it's clever: multi-agent systems with real approval gates, cloud automation that fails loudly, and fixes that arrive with the test that proves them.

Agent systems people can actually supervise. Most of my time goes to the parts that decide whether an AI system is safe to leave running: who is allowed to act, what happens when a step fails, and how a person stays in the loop without reading every log.

Upstream fixes, evidence first. Before I touch someone else's project I check that the issue is real, unclaimed and wanted, read the contribution rules, and write the failing test. Then the fix.

Cloud and defensive security. Container builds, AWS reliability work, threat modelling, and hardening the boring paths attackers like best.

• One agent talks to me. Forge holds my priorities and turns decisions into plain English. CLOUD runs the swarm and reports back to Forge in a fixed format, so nothing gets lost between them.
• Nothing leaves the machine without me. Pushes, PRs, posts and purchases wait for my approval of the exact item. An agent's confidence is not approval.
• Cheapest model that can do the job. The lead picks a model per task and escalates only after a real, revised attempt fails. A model fallback is logged, not treated as a blocker.
• One source, two platforms. A single registry generates the same agents for Claude Code and GitHub Copilot CLI, with tests that fail if they drift.

A local LLM writes a Dockerfile, docker build judges it, and the build errors feed the next attempt. A small loop with a ground-truth validator; documented samples cover Flask, Node.js and Go.

• career-ops #3711 — detect project-scoped Playwright MCP installs, with precedence-aware config and regression tests.
• career-ops #3591 — harden context-budget prompt assembly and tracker-path handling.

• mypy #22438 — treat writable callable attributes as invariant in PEP 695 variance inference.
• ruff #30317 — F402: report loop variables that shadow imports inside conditional branches, matching pyflakes (preview-gated).
• ruff #30318 — PLW0406: report from <module> import of the current module, matching pylint (preview-gated).
• conductor #1831 — keep the MySQL queue's unack sweep running after one failed run, so a dropped connection doesn't strand workflows until restart.
• agent-browser #2549 — let frame "#selector" resolve cross-origin iframes, with an end-to-end test against real Chrome.

• A bug report becomes a failing test before it becomes a diff
• "Works on my machine" is an environment finding, not a fix
• A mismatch between code and docs might be a design decision; I ask first
• Every external action is non-idempotent until proven otherwise: gate it, don't retry it
• Notes are part of the system: decisions and lessons get written down, dated, and corrected when they turn out wrong