AI in Engineering
Transparent about how AI fits into my engineering workflow. Where it helps, where it doesn't, and the guardrails that keep quality high.
I use AI agents across agency delivery, from triaging a client bug report through to the pull request that fixes it. This page covers which tools I use, what I've built with them, where they help, where they don't, and the guardrails that make them safe to run against client code.
Like any tool, the value comes from knowing when to reach for it and when not to. The guardrails matter as much as the capability.
What I've built
AI work delivered inside the agency.
AI in Practice
Posts on real AI usage: experiments, observations and what I took from them.
Operational Maturity Before AI Velocity: What Your Magento Team Needs First
AI is a force multiplier for Magento development — but it multiplies whatever processes you already have. Here's the operational stack your team needs before leaning heavily into AI-generated output.
AI Experiments for Frontend, SEO and UI Design
I rebuilt small projects to test AI for frontend, SEO and UI. Here’s what helped, what broke, and how I’d use it again.
How I use it responsibly
Every piece of AI-generated code or content goes through the same quality gates as everything else. The AI does not have a shortcut to production.
Sandboxed environments
All AI tooling is run within a sandboxed environment, preventing a rogue agent from affecting production systems or accessing sensitive data.
CI pipeline
Static analysis, automated test suites, and linting run on every change. AI-generated code that doesn't pass CI doesn't merge - no exceptions.
Human in the loop
Every AI-generated change is reviewed by a human prior to release. Shifting responsibility to the human reviewer, not the AI - the AI is a tool, not an authority.
No secrets or sensitive data
Nothing sensitive - always mount the bare minimum to the sandbox AI environments. Significantly reducing the scope of credentials and proprietary business logic available.
Verify don't trust
Testing is important, AI confidently hallucinates. Automated test suites, manual testing and code reviews are essential to ensure fixes and features function correctly.
Tools I use
Claude Code
My main coding agent, for implementation, debugging and review. Project instructions and skills keep it on house conventions rather than its own.
OpenAI Codex
A second coding agent alongside Claude Code, so a task is not tied to one model or one vendor.
BeeHive
My own agent orchestration tooling. Runs and keeps track of agent work across projects, so I am not babysitting a terminal per task.
MCP and CLI wrappers
Small MCP servers and CLI wrappers that give agents scoped access to analytics, Core Web Vitals and SEO data. They return validated fields, not raw API responses.
Where it genuinely helps
- Boilerplate-heavy work that follows known patterns: Magento admin grids, DI wiring, scaffolding
- Test generation for existing logic: produces a solid first draft to extend and verify
- Code review on pull requests: fast, consistent, catches things humans miss when tired
- Documentation from code: extracting intent and writing it clearly, faster than writing from scratch
- Exploring unfamiliar APIs and codebases: faster than reading docs cold
- Refactoring and making code consistent: pattern matching across large files
Where it still falls short
- Complex, stateful business logic: AI doesn't have the full domain context
- Security-sensitive code: never accept AI output without deep manual review here
- Long-term CSS and frontend maintenance: additive by nature, gets messy fast
- Reasoning about existing production systems it has not seen
- Anything requiring real-world operational judgment or incident experience
Using AI responsibly in engineering
Happy to talk through how AI fits into modern development workflows, compare notes on what's actually working in production, or discuss where the tooling still has gaps.