I build systems that keep working after the demo ends.

$ whoami
builder / operator / professional edge-case collector

$ current_status
teaching AI agents to bring receipts

I work where AI agents, production web systems, and automation meet. Most of my time goes into turning promising prototypes into boringly dependable software.

• Build agentic coding and operations systems with multi-agent delegation, skills, memory, browser automation, and scheduled jobs
• Work across open models and open tools instead of locking workflows to one provider or coding agent
• Turn search demand into useful products through technical SEO, programmatic SEO, structured data, and content pipelines
• Ship on Cloudflare Workers and OpenNext, then solve the less glamorous parts: cache correctness, cold starts, CI/CD, observability, and rollback
• Harden everything with tests, watchdogs, and live verification

I started with PHP and Drupal. These days, I spend more time making agents useful without letting them mark their own homework.

• Hermes Agent — working on agent reliability; my current contribution fixes duplicate compression notices and adds regression coverage (PR #71488)
• Plexus — fixed Codex SSE aggregation for non-streaming clients (merged PR #739)
• pi-qwencloud-token-plan-provider — my provider extension for using Qwen, GLM, and DeepSeek subscription models in pi
• Drupal.org — 19 credited issue contributions across 14 projects (full contribution record)
  • Elasticsearch Connector: fixed a PHP 7.4 array-offset notice in SearchBuilder::flattenKeys
  • Bibliography & Citation: patched missing <ref-type> data in EndNote8 exports
  • SMTP Authentication Support: removed deprecated magic-quotes code for modern PHP compatibility
  • RNG: fixed missing configuration export definitions for Drupal 9

bug reported
  -> reproduce it
  -> fix the cause
  -> add the test
  -> run the real path
  -> verify the result
  -> then say "done"

A few operating rules:

• Outcomes over ceremony.
• Every highlighted bug deserves a regression test.
• A green command is evidence, not proof that production works.
• If an automation says it succeeded, I go looking for the receipt.

Reliable long-running agents, model routing, context and memory systems, deterministic automation, browser reliability, edge performance, and making AI-generated work survive contact with reality.

_{Usually found somewhere between a failing test and a Cloudflare limit.}