AI Agent Architect · Sydney, Australia
I've been taking things apart since I was a kid. iPhones, Android systems, drones, motherboards. If something was a black box, I wanted to open it. That instinct never left.
I spent years building my own smart home from scratch—voice control, cross-platform integration, open-source components wired together over late nights. Not because I needed a smart home. Because I couldn't leave the problem alone.
From tinkering to building
In 2012, I moved to New Zealand and turned that curiosity into a career. My first real test was building the wholesale telecom platform that hundreds of ISPs depended on—fibre, broadband, mobile, all running through one system. When that platform went down, entire businesses stopped. I learned quickly that reliability isn't a feature. It's the whole point.
After that, I built systems that scaled: platforms serving hundreds of thousands of users, apps reaching millions. The architecture I designed years ago is still running today. I don't say that to impress—I say it because longevity in software is rare, and it taught me what actually matters.
Now
I'm a senior engineer at Iress. Outside work, I build ForgeLingo and BuildMate at Light Tune, my company, and take on a few privacy-sensitive AI projects. See the work →
What drives me
Everything I do comes back to making things.
I didn't plan a career. I followed what I couldn't stop thinking about. The jailbreaking led to the smart home. The smart home led to enterprise systems. The enterprise systems led to AI. Each thing I built opened a door I didn't know existed.
That's the thread: I create. Not for the output, but because the process itself feels like breathing. Code is just the current medium. Tomorrow it might be something else. The compulsion to make things—to understand how they work, to improve them, to ship them—that's the constant.
Two years in Sydney now. These days I take apart AI systems. Same instinct, different medium.
Things I believe
AI should reduce repetition, not replace judgment. The goal isn't fewer people—it's people doing more meaningful work.
Simple solutions beat clever ones. If a single API call solves your problem, don't build a multi-agent system.
Health is the infrastructure you can't outsource. I train daily—not because I love it, but because sleep quality affects code quality.
How I keep records
I log every AI-assisted working session I run, building or research, and the ones worth keeping get a retrospective: what I tried, what broke, what I changed my mind about. Since May 2026 that's over 3,400 sessions on one machine. Most are still raw logs or stubs; about 700 have a full retrospective. So I treat the records as raw material, not gospel.
For now the records are context: they show what I've been working on. Future pieces will name the sessions they draw from, separate what the record shows from what I remember, and mark anything I couldn't verify. Pieces published so far, including tutorials from 2025, don't carry that trail.
3,450 sessions, May–Oct 2026. 3,117 sorted into the topics above; 282 unsorted and 51 private are left out.
August 2026
Evals & verification
- Sessions
- 263
- With a full retrospective
- 8
Example sessions
- May 2026Checked a hallucination holdout segment by segment; caught a reversed legal condition and one unlabelled omission.
- Jun 2026Traced language-direction misjudgements to prompt bias rather than the parser; proposed a deterministic pre-check.
- Jun 2026Independent QA of an image-cache script: a bounded smoke test with checksums confirmed inputs stayed untouched.