// THE GEEK PAGE

How this thing
actually works.

A proper, no-fluff, behind-the-scenes look at how Draxiq builds you a complete bespoke website in about 60 seconds — and the same technology that powers our chat assistants, Telegram bots and custom automation. If you like servers, APIs and the smell of fresh terminal output in the morning, you'll enjoy this.

01 The Stack

Let's start with what's actually running. Draxiq lives on a DigitalOcean droplet in New York (yes, we should probably move it to London — it's on the to-do list). It's a humble little 1GB Linux box running Ubuntu 24.04, and despite its modest specs it punches well above its weight.

🐧

Ubuntu 24.04 LTS

The OS doing all the heavy lifting. Stable, fast, free.

Nginx

The reverse proxy out front, handling SSL and routing traffic.

🟢

Node.js + Express

The actual app — handles chat, generation, leads, the lot.

🔄

PM2

Process manager. Keeps everything running 24/7, restarts on crash.

🔒

Let's Encrypt SSL

That little padlock in your browser. Free, auto-renews every 90 days.

📨

Brevo

How leads get from the website into Graham's inbox.

It's a deliberately boring, deliberately battle-tested stack. No exotic frameworks. No things that'll be deprecated next Tuesday. Just the stuff that works.

02 Meet Ivy

Ivy is the friendly face you chat with on the homepage. She's not a chatbot in the old "press 1 for sales" sense — she's a proper conversational AI built on top of Anthropic's Claude Opus 4.6, the most advanced AI model on the planet right now.

Under the hood, every message you send Ivy gets bundled with a carefully written system prompt — basically a long set of instructions that tells her who she is, how she should sound (warm, Southport-friendly, plain English), what she's trying to find out (your business name, what you do, the vibe you want, your colours, your services), and crucially, when she has enough info to start building. She also handles enquiries about chat assistants, Telegram bots and custom automation — so visitors interested in our other services get a friendly chat too.

// Fun fact Ivy is told to never overwhelm you with questions. She asks one thing at a time, conversationally, like a human would. The temptation when building these things is to make a giant form. Forms are boring. Conversations aren't.

When Ivy decides she's gathered enough, she does something sneaky: she outputs a special tag on a new line — [GENERATE] followed by a JSON object containing everything she's learned. The frontend sees that tag, hides it from you, and fires off the actual website build.

03 The Build

This is where it gets fun. The build endpoint takes Ivy's gathered details and sends them to Claude Opus 4.6 with a much bigger, much more demanding system prompt — one that says, in essence, "build me a complete production-quality multi-page website right now, output the HTML directly, no preamble, no markdown fences, just shut up and design."

But here's where it gets clever. Left to its own devices, Claude has a "house style" it tends to fall back on — sticky nav, hero section, services cards, gallery, testimonials. Beautiful but samey. So we built a variant lockdown system.

// Every build, the server randomly picks ONE of these: const STYLE_LOCKS = [ 'BRUTALIST', 'EDITORIAL_MAGAZINE', 'DARK_GLASSMORPHISM', 'RETRO_90S', 'SWISS_MINIMAL', 'NEO_PLAYFUL', 'ORGANIC_NATURAL', 'BOLD_GRADIENT', 'LUXURY_GOLD', 'TERMINAL_HACKER', 'NEWSPAPER', 'CUT_AND_PASTE_ZINE' ];

Each style comes with concrete, non-negotiable rules. Brutalist gets thick black borders, square corners, monospace fonts and a single neon accent. Editorial Magazine gets cream backgrounds, drop caps, and refined serifs. Terminal Hacker is pure black with phosphor green. We also randomise the hero layout (8 variants) and the section order (7 recipes), so structurally no two previews are ever the same.

The system prompt then begs Claude to commit. Literally. There's a line in there that says "if your first instinct is sticky-nav-with-rounded-cards-and-soft-shadows, STOP and reread the style rules." And it works.

04 The Images

A website without images is just a CV. But generating six unique, high-quality, on-brand images in real time is tricky business. We tried a few options before landing on the right one.

The early attempts. First we tried Pollinations (free, but slow and unreliable — kept timing out). Then we considered DALL-E 3 (gorgeous but slow and pricey). Then Runway (great for video, not its strength for images, and expensive). The winner, by a country mile, was fal.ai's Flux Schnell.

fal.ai Flux Schnell

Generates a 1536×1024 image in about 2-4 seconds. All six images for a preview generate in parallel — total image build time is under 5 seconds.

🎨

Pixar-quality cartoon

Every prompt is auto-suffixed with style instructions — "vibrant flat illustration, Pixar quality cartoon, bold colors, crisp clean lines, cheerful."

But here's the really clever bit. We don't trust Claude to write image URLs. Claude is brilliant but it's been trained on a billion Unsplash links and will reach for them on instinct. So instead, Claude is told to output {{IMG:description}} placeholders, and the server intercepts those and replaces them with real fal.ai URLs after generation. Mechanically guaranteed. Bulletproof.

// Cost transparency Each preview costs about 85p in API calls — roughly 80p for the Claude generation and 5p for all six fal.ai images combined. We don't charge you to try Ivy because we want you to be impressed enough to buy the real thing afterwards. Yes, this means people can technically waste our money. We've decided we're OK with that.

05 Speed Tricks

A 1GB droplet running multiple Node apps shouldn't be this responsive. So we cheat. Politely.

Gzip compression on every response shrinks pages by about 70% over the wire. Browser caching on static assets means after your first visit, refreshes are basically instant — your browser doesn't bother re-downloading anything. Image pre-warming fires off all the fal.ai requests in parallel the moment Claude finishes generating, so by the time you click "Open Preview" the images are already cached. Shared CSS files mean once you've loaded one Draxiq page, every other page on the site loads almost instantly because the styles are already in your browser cache.

Add it all up and you get a site that feels snappy on a server that costs less than a pint of beer per month.

06 Security

The basics are all here. Full HTTPS via Let's Encrypt, auto-renewing every 90 days. Rate limiting coming soon (so nobody can spam-generate 50 previews and burn through credits). And every line of frontend JavaScript gets auto-obfuscated on server startup using javascript-obfuscator with control flow flattening and string array encoding — so if you "view source" you get a wall of unreadable hexadecimal gibberish. It's not bulletproof but it stops 99% of casual copy-pasters dead.

07 The Live Numbers

Because we love a good stat. These are real and current.

12
Style variants
8
Hero layouts
~60s
Build time
85p
Cost per preview

08 Why We Built It This Way

Because most "AI website builders" are dreadful. They use templates. They give you a drag-and-drop editor and call it AI. They charge you £30 a month forever for something that looks like every other site on the internet.

We wanted something different: a real conversation, a real bespoke design every time, a real working preview you can actually scroll, and a one-off honest price (£200 for small businesses, £350 for bigger ones). No subscriptions. No nonsense. No stock-photo soup.

And the same thinking applies to everything else Draxiq does — chat assistants, Telegram bots, custom automation, dashboards. Frontier AI used properly, by humans who care, at honest prices.

And we wanted to build it in Southport, for Southport — and then for everyone else who deserves better than another Wix site.

09 What's Next

Plenty. The roadmap includes: a "recent creations" gallery on the homepage, automated SEO setup baked into every site we deliver, AI-generated blog posts running on autopilot, a Southport-specific puzzle that wins the solver a free site, expanding our chat assistant offering, more dedicated service pages, and eventually opening Draxiq up to teams of designers and developers who want to build with frontier AI. One step at a time though.

// One small acknowledgement The brain inside Ivy — the actual writing, designing and decision-making — is powered by Claude Opus 4.6 by Anthropic. The current state of the art in large language models, and frankly the only model good enough to do what we're asking it to do. None of this would be possible without it. Respect where it's due.

Right then. Want to see it in action?

Stop reading. Go and chat with Ivy. She's been waiting.

Talk to Ivy →