Some of the best architectural decisions I've made never went to production. This is one of them.
The Breaking Point
We were running a high-traffic marketing site - heavy A/B testing, dynamic banners, offers, primarily organic traffic from email opens. The kind of site where first impressions are everything and every second of load time has a direct business cost.
Our stack was Gatsby. And it was falling apart.
Build times were getting out of hand. Load times were degrading. Then Netlify acquired Gatsby and support effectively stalled. We also started hitting Netlify downtime and third-party integration blockers that compounded an already fragile setup. The site was expensive to maintain and getting slower.
Leadership assigned me to research a new stack. The brief was open - evaluate options, build POCs, present findings.
The Room Had Already Decided
The moment it was brought up in standup, the team rallied around Next.js. Everyone had experience with it. Everyone had opinions on why it would work. It was the obvious choice.
I understood why. Next.js is a mature, well-supported framework with a strong ecosystem. For most use cases, it's a reasonable default.
But I wasn't there to validate consensus. I was there to find the right answer for the product.
So I ran the Next.js POC anyway - properly, without shortcuts. The developer experience was good. The ecosystem was solid. And then I looked at the output.
Non-interactive components were being hydrated. Common JS bundles were being shipped to the client unnecessarily. The thread blocking time was higher than it needed to be for a site where most components never respond to user input. For an application, this is an acceptable tradeoff. For a marketing site built to load fast and convert, it's the wrong architecture.
The framework was technically sound. It was the wrong tool for what the product actually needed.
Finding Astro
I kept researching. Astro had been gaining traction with a specific promise: ship zero JS by default, hydrate only what needs to be interactive. For a marketing site, that's not a feature - that's the entire architectural requirement.
I built the POC. Same content, same structure, equivalent functionality.
The numbers were not close.
| Metric | Gatsby (Production) | Next.js POC | Astro POC |
|---|---|---|---|
| FCP | 2s | ~1.2s | 0.2s |
| LCP | 10s | ~5s | 2s |
FCP dropped 90%. LCP dropped 80%. These weren't micro-optimizations - they were the direct result of not shipping unnecessary JavaScript to a page that didn't need it.
Where AI Fit In
I used AI heavily during the POC phase - wireframes, component scaffolding, boilerplate generation. It compressed weeks of setup into days.
But AI didn't make the architectural call. It couldn't reason about why a marketing site with non-interactive components has no business hydrating them. It couldn't weigh the tradeoff between developer familiarity and production performance. It couldn't recognize that the team's instinct toward Next.js was a comfort zone decision, not a product decision.
That judgment of understanding what the product actually needed and selecting the tool that served it. That's where the human work was.
What Happened After
I presented both POCs to leadership with full benchmarks. The Astro numbers were clear. The recommendation was clear.
Then the whole team was laid off. Company was on the brink of an IPO, circumstances changed, and that was that. Neither stack ever went to production.
What I Walked Away With
The work didn't ship. The thinking did.
When you're assigned a technical decision, the room will often have already made up its mind. Your job isn't to confirm that. Your job is to understand what the product needs and find the architecture that serves it - regardless of what's familiar, regardless of what's popular, regardless of what the consensus is pointing at.
AI will keep getting better at execution. It will scaffold faster, generate more, automate more. What it won't do is walk into a standup, hear a team default to the comfortable answer, and ask whether that answer is actually right.
That's still the job.
One More Thing
Recently, Cloudflare acquired Astro. A framework built to ship minimal JavaScript, backed by one of the largest infrastructure companies on the internet.
When I ran those POCs, Astro was the underdog pick against a category default. That acquisition is validation, not just of the framework, but of the architectural reasoning behind choosing it.
The right tool for the product, before the market agreed.
Top comments (0)