21 → 22 June 2026 repo to live in ~1 day
Solar Pays Off: 160+ pages in a day, honestly
solarpaysoff.com
A deliberate revenue experiment: three US energy calculators (solar payback, EV charging, AC running cost) rendered per-state for all 50 states plus DC, funded by display ads. Repo created June 21; on my portfolio as a live site June 22.
- Stack: Astro static build, one data engine feeding 160+ pages from EIA electricity rates, NREL sun data and the federal credit rules.
- What AI did: the entire first cut — data pipeline, page templates, calculator logic. My job was making the numbers honest and killing thin-content pages.
- The judgment call: programmatic SEO has a deserved bad name. Every state page had to earn its place with genuinely different data, or the whole thing is spam.
The lesson: AI makes the build nearly free. The scarce ingredient is knowing what's worth building and where the honesty line sits.
27 → 28 June 2026 repo to live in ~2 days
EA Centre of Excellence: 27 modules in a weekend
ea.rexe.au
A free enterprise-architecture learning site — 27 modules across five rings, seven quizzes, a worked case study and an exportable capstone. Built in about two days, which sounds impossible until you understand the division of labour.
- What AI did: drafted every module fast, structured the pathways, built the Astro site and quiz engine.
- What I did: the part AI can't — 35 years of watching EA fail in practice. The opinions, the anti-patterns, the "here's what the framework courses won't tell you" judgment calls. AI typed; the scar tissue is mine.
The lesson: AI collapses the cost of expressing expertise. It doesn't create the expertise. That's exactly the leverage: decades of judgment, shipped at weekend speed.
Late June → early July 2026 live and taking payments
DesignMyInk: a real commercial AI SaaS, end-to-end
designmyink.com
The full commercial test: an AI tattoo-design generator with real image-generation costs, real billing, real customers. Describe your tattoo, get studio-ready flash in ~30 seconds in any of 60+ genuine styles.
- Stack: a single Cloudflare Worker running the whole product — frontend, generation pipeline, freemium billing.
- What bit me: Workers CPU limits. An image-heavy product on a platform with hard per-request CPU budgets forces you to be surgical about what runs where.
- The product work: 60+ styles is a curation job, not a prompt job — each style has to look like something a tattooist would actually recognise.
The lesson: "AI product" usually means a demo. Billing, quotas, abuse handling and unit economics are where it becomes a business — and that's most of the work.
15 June → early July 2026 ~3 weeks of iteration
SeparationSense: modelling the maths nobody shows you
separationsense.com.au
The Australian child support formula is eight steps of opaque maths, and it lands on people in the worst months of their lives. This models both households' weekly cashflow — child support, Family Tax Benefit, Parenting Payment, Rent Assistance — entirely in the browser.
- The hard part: the interactions. Child support counts against FTB through the Maintenance Income Test; payments move when income moves. No single government tool shows the combined picture — which is precisely why this exists.
- The deliberate constraints: local-first (nothing leaves the device), estimate-only, and loudly not-advice. Getting the disclaimers and the crisis-safety right took real care.
- What AI did: implemented and cross-checked the formula logic against the published rules while I attacked it with edge-case scenarios.
The lesson: deterministic, testable domain logic is where AI-assisted coding shines — every rule has a published answer to verify against. The design judgment is deciding what not to promise.
3 → 21 June 2026 ~2.5 weeks
notNPD: writing calm content for an angry topic
notnpd.com
A companion to mentalhealth.men. "Narcissist" has become the internet's default word for any difficult man — and the people it lands on have nowhere calm to think it through. Everything online either damns them or absolves them. This does neither.
- The editorial tightrope: hold two truths at once — the label is wildly overused, and real narcissism exists. Every page got rewritten until it stopped taking a side.
- The engineering that matters: the anonymous reflection tool keeps no score and sends nothing. Not a privacy policy promise — an architecture decision.
- Crisis-safe throughout: non-diagnostic language, real Australian help pathways, Lifeline everywhere it should be.
The lesson: on sensitive topics the code is the easy 20%. The other 80% is tone — and AI drafts tone fast, but a human has to own it.
Feb 2026 → ongoing the slow-burn one
City to Surf: rescuing 320,000 results from old newsprint
city2surf.rexe.info
The early results of the world's biggest fun run — from 1,500 finishers in 1971 to tens of thousands — existed only in publicly published media of the era. Effectively lost. Now they're a searchable archive: every runner, every finish time.
- The grind: digitising era print media is messy — OCR errors, name variants, inconsistent formats year to year. Most of the project is data cleaning, not web dev.
- What AI did: the extraction and reconciliation pipelines — and tireless pattern-matching on mangled OCR that would have broken a human transcriber's spirit.
- Why bother: because someone's dad ran it in 1974, and now they can find him.
The lesson: not everything is a one-day ship. Heritage data rewards patience — and it's the project strangers email me about most.
The pattern across all of them
The division of labour is consistent. AI — almost always Claude Code — does the heavy lifting: first drafts, boilerplate, pipelines, refactors, the 3am-grind work. I do the parts that come from 35 years of shipping: choosing what to build, where the honesty line sits, what the edge cases are, when it's actually done, and putting my name on it.
Speed is real, but it isn't the point. A day-to-live site and a five-month data rescue sit on the same list. The point is that one person can now carry an idea from "that's a real problem" to a live, maintained product — and do it twelve times.
Want something built like this?
Idea to live product, fixed problem, or a second opinion — tell me what you're trying to do.
Start a conversation