Work

What it is

A marketplace where homeowners list privately and verified estate agents subscribe by area to see those listings. It runs the whole business, not a prototype: payments, email, consent records, an admin review queue, SEO area pages and advert tracking.

How it was built

Two AI tools, each for what it does best. An app builder, Lovable, generated the first version from prompts. Then Claude Code worked as the senior engineer under rules I wrote: question every decision, back anything touching money or personal data with code or a test, write every decision down, and review one named area at a time, on a schedule.

Built to a professional, production-ready standard

The generated code worked, but working isn't the same as ready to launch. Scheduled reviews brought it, area by area, to the standard I'd expect of any production system.

• Security: Access rules are enforced in the database and checked again on the server, so no one can see or change what they shouldn't. Prices are set on the server, never in the browser.
• GDPR: Tracking runs only behind consent, consent is recorded with the exact wording shown, and personal data is kept away from ad platforms.
• DevOps: Staging and production are separate, each with its own database and payment account. Every release passes a smoke test of the site, every server function and every third-party service, then ships through a reviewed pull request.
• System alerting: Errors are reported to an admin page, admins are alerted when listings or agents are waiting for review, and every incident since launch has become a new check.
• Scalability: Managed hosting and database grow with traffic without code changes. A new city is mostly data and content, not new code, and the written rules let a second developer, human or AI, work to the same standard from day one.
• Cost management: There are no servers to run. Public pages are static and served from a CDN, so a marketing spike costs little and never touches the database.

Automated marketing, with a person approving every post

A team of Claude agents runs the social schedule. They do the legwork; nothing goes out until I approve it.

• News scouting: Twice each weekday an agent reads UK property news, scores each story against set criteria and drafts only the strongest, with copy for LinkedIn, Facebook and Instagram and an image checked against strict rules. Of 251 stories assessed so far, 10 were drafted.
• Approval: Every draft goes to a Teams channel and waits for my "approve", "edit" or "skip". The agents accept an approval only from me, and treat every other message as information, never as an instruction.
• Publishing: Approved posts are scheduled for each network's best time and published to the site's news page, tagged so a sign-up can be traced back to the post that brought it.
• Guardrails: Copy may use only claims from an approved list, every statistic carries its source, and there is no hype, urgency or emoji.
• Learning: A weekly review reads the analytics and suggests changes. The ones I approve go into a file every agent reads on its next run.
• Google Ads: Set up privacy first: Google's tags load only after consent, and the sign-up conversion carries no personal data. Enhanced conversions were turned down because they send a hashed email address, and the same rule applies to Meta.

What it shows

AI is only as good as the person driving it. It writes working software quickly, but it doesn't know what to check, and in inexperienced hands that ends in a launch nobody can trust. Knowing what to review, test and write down comes from experience. That's the part I bring to a build.

What it is

Magical Mushroom Company grows packaging from mycelium as an alternative to plastic foam. The online business sold its off-the-shelf range, from gift boxes and bottle cases to protective corners, with quantity discounts, pre-orders and Royal Mail shipping built in.

How it was built

Headless, with each part chosen for its job: BigCommerce for commerce, Storyblok for content, Next.js on Vercel for the storefront and Stripe for payments. I'd moved a large luxury UK retailer onto BigCommerce before, so I knew how far it could scale. A design agency, a development agency and a small in-house team worked to one plan, reviewed with me every week. The shop launched inside the main website first, then moved to its own brand and domain. It ran with near-zero downtime, atomic deployments with automated testing and shared components, at a fraction of the cost of a comparable enterprise build, and was featured by The Headless Club.

Run as part of the whole business

The shop was one part of a wider CTO remit, run to the same standard as the rest of the company.

• Cost management: Platform costs were planned against projected sales, so each tier was taken on only when the business needed it.
• Security: Two-factor sign-in enforced, software bought as a service rather than built, and company laptops centrally managed.
• GDPR: I was the Data Protection Officer: registered with the ICO, with a written data-breach policy and an IT and communications policy for all staff.
• Resilience: A disaster recovery plan with a calling tree, and a backup review of each core system.
• Systems: A Monday.com workflow from first enquiry to production, and the selection of a manufacturing planning system.

What it shows

A small company can use the same architecture as a large retailer if it pays only for the capacity it needs. The skill is knowing which enterprise principles to keep from day one, and which costs can wait. That's what I bring to a build.