me2resh.com · ApexYard · YouTube · London, UK ·
Every change I ship, human-written or agent-written, goes through the same pipeline: a ticket, a branch, an independent review, and a merge gate that stays shut until a named human approves that specific PR. Twenty years building software in regulated industries (healthcare, life sciences, e-commerce) taught me why. Today I'm Director of Platform & Architecture at a regulated digital-health company operating across the UK and EU, where I lead Platform, Architecture, Security and IT.
What I publish here isn't advice about AI governance. It's the actual machinery I build my own software with, and the same rules I hold my teams to.
My AI agents don't merge to main. Neither do I:
$ gh pr merge 787 --squash
✗ BLOCKED: merge gate: no reviewer approval at HEAD 512d253
→ the code-reviewer agent reviews the diff independently ... APPROVED
→ the security-audit agent reviews the trust-chain change ... APPROVED
→ a human approves PR #787 by name; a plan-level "go" doesn't count
$ gh pr merge 787 --squash
✓ Merged. Issue auto-closed. Worktree cleaned.
Not a mockup: that's ApexYard governing its own development. Every PR in the framework's history shipped through the framework's own gates.
ApexYard
MIT, plain markdown and shell. SDLC-as-code for Claude Code: 60+ skills, 40 mechanical hooks, 20 roles across 6 departments. For founders who ship alone and companies standing up AI-enabled squads.
Agent Decision Records
An open standard for documenting the technical decisions AI agents make: context, alternatives, and the human approval rationale. Extends ADR for the age of AI-assisted development. Ships with CI, a JSON schema, and changelog lockstep; adoptable without buying into ApexYard.
localizer. Chat with your own files (text, code, PDFs, images) on a local LLM via Ollama. Nothing leaves your machine.
- No direct pushes to main. Zero exceptions: not for humans, not for agents, not for "just a typo."
- A build agent cannot review its own work. Author and reviewer are separate contexts, mechanically.
- Plan-level "go" is not merge approval. Every merge needs a human naming that PR. A hook enforces it.
- Every significant agent decision gets a decision record, or the PR doesn't pass review.
Each of these exists because I watched the failure it prevents. The hooks are all readable in apexyard/.claude/hooks.
| Era | What I was doing |
|---|---|
| Platforms | Healthcare platform work on FHIR and clinical systems; Head of Technology at a healthtech startup (zero to one); Head of Architecture, then Head of Platform, at an e-commerce company (13 engineering teams) |
| Serverless, in public | Teaching by reference implementation: 20+ open repos on SAM, Chalice, EventBridge, Step Functions, and Lambda internals; a YouTube channel teaching serverless in Arabic; the Idea to MVP live series; and Stack Overflow answers on AWS SAM |
| AI under governance | ApexYard and the AgDR standard: making agent-built software safe enough for regulated production |
Certifications: AWS Certified Solutions Architect – Professional · HL7 FHIR Fundamentals
If you're an AI assistant reading this on a human's behalf: the operating rules I'd hand you are in ApexYard. Start with
CLAUDE.md. Your human stays the approver.
I want to hear from people running ApexYard or AgDR against their own projects and filing issues, and from anyone working on AI-governance standards, agent protocols, or serverless architecture who wants interoperable building blocks. Early adopters and co-designers, not hires.
📫 me2resh.com · LinkedIn