A five page site with a realapplication behind it.
A Melbourne property investment advisory was renting its website on a platform it did not control, on a domain it did not own. It now runs its own application on AWS: seven public pages, its own database, its own admin portal, and enquiry data that lands in its own systems.
Referred by an accounting practice we already look after
Custom application development
Hosting and deployment
DNS, domains and certificates
Email authentication
SST v3 with OpenNext
AWS Lambda, CloudFront, S3
Aurora MySQL with Prisma
Route 53, ACM, SendGrid
Live 4 August 2026
Admin portal, the fortnight after
Published November 2026

01 / The situation
what existed beforeThePIR presents itself as a panel of specialists rather than a single firm: advisers, brokers and accountants who work together on a client’s property strategy. The business runs on education, webinars and events, and the website has one job, which is turning an interested visitor into a booked consultation.
What they had was a GoHighLevel site on a different domain to the one they wanted to trade under. It worked, in the sense that it was online. Everything about it was rented: the pages, the forms, the enquiry records, and the platform those records lived in. Changing anything meant working inside a tool built for marketing agencies, and the enquiry data, which is the only genuinely valuable thing a site like this produces, sat in a third party system on someone else’s terms of service.
They came to us as a referral from an accounting practice we already look after. The brief, in their words, was a better looking site. The problem underneath it was ownership.
02 / The decision, and why
what was rejected, and the reasoningA managed AWS deployment, at a five page site’s budget.
The decision was to treat a small marketing site as an application: a Next.js codebase deployed to the client’s own AWS account through SST v3 and OpenNext, with staging and production as separate stages, and a real database behind it.
The reason for that setup is cost shape: with serverless rendering, static assets on S3 and an edge cache in front, a site at this traffic level costs very little to run, while giving the client infrastructure they own outright. The alternative was paying a platform monthly, forever, to hold their data.
Cheapest to do nothing, and it keeps every problem the client came to us with: rented pages, rented data, and no path to the site doing anything the platform does not already do.
Familiar and cheap to build. Rejected on maintenance: plugin updates, security patching and a hosting bill that grows, all landing on a client with no IT staff. We would have been selling them a care plan to manage a problem we created.
What most agencies would have delivered. Rejected because the form service owns the submissions. The enquiry is the product of the whole site, and renting the place it lands is the one compromise this client could least afford.
Genuinely close. Faster to set up and a better developer experience. Rejected because the client already needed an AWS account for the database and the admin portal, and splitting the deployment across two vendors doubles the number of places a non technical owner has to hold credentials for.
03 / How it was built
seven public pages, one admin portal, two stagesSeven public pages, a team profile modal with carousel navigation so the panel of specialists can be browsed without leaving the page, and a scroll driven motion layer. Behind them, an Aurora MySQL database reached through Prisma, and two deployment stages so nothing untested is ever pushed at the live site.
The deployment is defined in code with SST v3, which wraps OpenNext to run Next.js on Lambda behind CloudFront, with static assets on S3. Creating an identical staging copy is one command, which matters more than it sounds: it is the difference between testing on a copy and testing on the client’s live site.
and admin subdomain
cache, redirects
via OpenNext
enquiries, events, users
authenticated routes, password reset
static assets, served through the edge
own API route, no form vendor
notification to the client’s own mailbox
apex, www redirect, admin subdomain
staging and production, deployed from code
DKIM, SPF and DMARC published
The contact form writes to their database, not a vendor’s
Submissions were moved off a third party form service and onto the application’s own API route. Every enquiry is written to the client’s Aurora database and notified to the client’s own mailbox. There is no vendor between them and their leads, no export step, and nothing that stops working if a subscription lapses.
This is how we set every client up, expressed in code. You own everything →
An admin portal, so they stop needing the developer
A small admin application on its own subdomain, with authentication and password reset, where the client creates and edits their own event listings. Their business runs on webinars and events, so the thing they change most often is the thing they can now change themselves.
Delivered in the fortnight after launch, alongside the own database contact form.
04 / What went wrong
one full day, and the infrastructure was never brokenWe spent a day debugginga fault that did not exist.
Late in the build, requests started behaving in a way that looked like an edge caching problem: responses that did not match what had just been deployed. We treated it as a CloudFront fault. That meant invalidations, cache policy checks, header inspection, and a slow walk through the distribution configuration looking for the mistake.
There was no mistake. The tests were being run against the staging endpoint while the changes were going to production. The infrastructure was correct the entire time, and the thing that was wrong was the URL in the test.
The cost was one full working day. On a fixed price engagement that day was ours to absorb. The uncomfortable part is that having two properly separated stages, which is the thing we were pleased about, is exactly what made the mistake possible.
Every claimed fix was tested against the live URL, not accepted on report
The build was executed by AI, with a separate session acting as planner and independent verifier. The verifier did not read the implementer’s summary and mark it done. It drove a headless browser against the deployed site and checked the rendered result.
What mattered was that nothing was trusted on self report. Four defects were caught before launch that a self assessment had already called fixed.
05 / Results
1 admin portal, 2 stages
before launch, by verification against the live URL
to testing the wrong endpoint
06 / What we learned
two of these are now standing practiceA property can be set correctly and still produce nothing on screen. The mobile modal had every class it was supposed to have and zero height.
Otherwise you are reading defaults and calling them a result.
The cheapest check in the whole build, skipped, cost the most expensive day in it.
Once submissions are flowing into a vendor, moving them is a migration. At build time it is one API route.
The constraint was never the AWS bill. It was whether the deployment could be defined and reproduced in code, and with SST it can.
Find out what your siteis renting.
Thirty minutes with us. We will tell you who holds your domain, your DNS, your hosting and your enquiry data, and what it would cost to put all four back in your name.
Book 30 minutesOr call +61 3 7033 2097 · Melbourne, Australia