Fifteen years of making operational systems understandable — and now building them. I direct AI coding agents, personally review and merge every change, and run the CI that checks them. Every stratum below is load-bearing.
I.
The record
counted from GitHub, not remembered
5,300+
pull requests I personally reviewed and merged
past 12 months · two GitHub accounts · GitHub-counted 15 Oct 2026
11
self-hosted CI runners I built and operate, all online
Linux x64 · Linux ARM64 · macOS ARM64 · registered 15 Oct 2026
14+
repositories with merged work
past 12 months · lower bound from a GitHub search sample
≈1M
support tickets migrated in production
approximate · led and coordinated, cross-functional
I direct AI coding agents: they draft the code; I scope each change, review every diff, and merge each pull request myself. In a 100-PR sample, 99 were merged by hand with auto-merge off.
II.
Skills, with receipts
each links to the plate that proves it
Directing AI coding agents
Breaking work into small, reviewable pull requests, writing the specs agents build from, reviewing every diff, and merging by hand — inside a workflow where status is checked, never inferred.
GitHub Actions on my own hardware: custom Docker runner images, label-routed lanes, cgroup resource envelopes, canary checks for runner identity, SHA-pinned actions, and a hardened boundary for untrusted pull requests.
Python, SQLite, and FTS5 search tools that run offline, and a re-profiled read path that cut a status payload roughly eightfold without dropping audit detail.
Led a Zendesk-to-HubSpot migration of approximately one million support tickets — cleanup, validation, workflow changes, QA, and cutover support, coordinated across teams while production kept running.
SystemsData migrationOperations
moved: ≈1,000,000 support tickets (approximate, qualified)
Designed and operates a governed workflow platform with explicit task states, leases, approval gates, human/AI handoffs, and verification artifacts — so status is checked, never inferred from prose.
Personal · AI-assistedEngineeringGovernance
surface: task states · readiness · assignments · leases
Heterogeneous self-hosted runners with canary checks for identity and cgroup envelopes, a custom Docker runner image, SHA-pinned actions, and a hardened untrusted-PR boundary — thousands of pull requests through it.
Personal · AI-assistedDevopsCI security
fleet: Dell linux-x64 docker-class + Mac mini runners
Authored documentation standards and implemented content models, metadata, permissions, review, QA, and publishing workflows for a production help center — so customer guidance stays maintainable without unwritten knowledge.
Roadmap and production-health ownership of a customer-facing knowledge platform, with hands-on application, testing, release-readiness, and troubleshooting work alongside collaborators.
A Python, SQLite, and FTS5 helper that indexes curated evidence and returns source-linked results with no network dependency — built because real-time recall is hard, and verified in implementation.
Re-profiled a live registry read path and cut the default status payload roughly eightfold while preserving full audit detail as an explicit option — with regression coverage to keep it honest.
An evidence-first family-history archive: source records modeled before narrative, record-level privacy decisions, and static publishing one person can maintain part-time.
PersonalKnowledge systemsInformation architecture
claims: every published statement traces to a named source record
privacy: publish or withhold decided record by record
maintenance: one person, part-time, static output
IV.
Method
Claims on this site keep their qualifiers. Counts marked ≈ are approximate and recorded as such; metrics I can't source don't appear at all. Personal projects are labeled, AI-assisted work is disclosed, and nothing here implies sole credit for cross-functional work.
That isn't modesty — it's the discipline the work is about.