01
Build from a brief
Build an AI agent that can create a complete website from a plain-language brief — or rebuild an existing one.
Case study
Articode
How Huboxt designed and built articode.ai from scratch — a configurable AI agent on Pydantic AI and FastAPI, wrapped in a full platform that lets anyone describe a website in plain words, publish it on a subdomain, and connect their own domain.
- Client
- Articode
- Timeline
- 3–6 months
- Services
- AI agent engineering, full-stack development
- Engagement
- Greenfield product
Product preview
The challenge
"Describe your website and we'll build it" is an easy promise to make and a hard one to keep. Doing it well means an AI that doesn't just generate a one-off mockup, but reliably produces a complete, coherent, publishable site — and a platform robust enough to host it, give it an address, and let the owner keep editing it by simply talking to it.
Articode came to the table with the vision and no codebase. Everything had to be built from scratch: the agent that does the thinking, the platform users log into, the publishing and hosting layer, custom-domain support, and a billing model that worked for an AI product where cost scales with usage. On top of that, the agent couldn't be wired to a single model — it had to stay flexible across the fast-moving landscape of public LLMs, so the product wouldn't be locked to one vendor's pricing, availability or capabilities.
The hard part wasn't any single feature. It was building an agent architecture flexible enough to evolve as fast as the AI space does, while shipping a polished, paying product around it.
Goals
02
Configurable agent
Make the agent configurable, not hard-coded: prompts, skills, build flows and model choice all adjustable without rewriting the system.
03
Model-agnostic
Stay model-agnostic across all the popular public LLMs.
04
Full platform
Deliver the full surrounding platform: accounts, publishing, hosting on subdomains, custom domains and billing.
05
Launch-ready product
Ship a real, launch-ready product — not a demo.
Our approach
We built the agent as configurable infrastructure on Pydantic AI, hosted it as a FastAPI service, integrated the major public LLMs from day one, and delivered the full platform in clear milestone-based phases.
01
Configurable infrastructure
Prompts, skills, the website-building flow, chat interactions and which LLM to use are all tunable configuration — not buried in code that has to be rewritten every time the product evolves or a better model appears.
02
FastAPI service boundary
The agent runs as a clean FastAPI service, giving the platform a fast, typed, well-defined boundary and a sensible place to scale the AI workload independently from the rest of the product.
03
Model-agnostic from day one
Integrated the major public LLMs and made model selection a configurable choice — so the product can route to the right model for the right job and adapt as the landscape shifts.
04
Milestone-based phases
Clear milestones — the agent core, then the platform and publishing layer, then custom domains and billing — so each phase delivered a real, working capability.
What we built
A complete AI product — configurable agent, user platform, multi-tenant publishing, custom domains, and usage-based billing.
- 01
The AI agent
A website-building agent on Pydantic AI, hosted as a FastAPI service, with two modes: build a brand-new site from a plain-language description, or rebuild an existing one. Prompts, skills, build flow, chat behaviour and LLM usage are all configurable, integrated with the major public LLM providers.
- 02
The platform
A full product around the agent: user accounts and login, a building experience where users describe what they want and watch the site come together, and one-click publishing.
- 03
Publishing & hosting
Multi-tenant publishing that puts every user's site live on its own *.articode.ai subdomain — the hosting and routing layer that turns a generated site into a real, shareable address.
- 04
Custom domains & billing
The ability to buy and connect a custom domain directly through the platform, plus a token-based billing model that charges users for the LLM usage they actually consume.
Results
01
Built and launched
A complete AI product, built from scratch and launched — agent, platform, hosting and billing — within a roughly 3–6 month engagement.
02
Live and in use
Articode is in the hands of real users building and publishing real websites in minutes.
03
Agent that ages well
Because prompts, skills, flows and models are configuration rather than hard-coded logic, the product can adopt new models and capabilities without re-engineering its core.
04
No vendor lock-in
Model-agnostic integration across the major public LLMs keeps the product flexible on cost, quality and availability.
05
End-to-end ownership
From "describe my site" to a published page on a subdomain, to a custom domain and usage-based billing, all in one platform.
“The win here was architectural restraint. It would have been faster, in the short term, to hard-code the agent to one model and one fixed building flow. Instead, Huboxt built the agent as configurable infrastructure on Pydantic AI — prompts, skills, flows and models all adjustable — and wrapped it in a complete platform that handles the unglamorous but essential work: accounts, publishing, hosting, domains and billing. The result isn't just a clever demo of AI website generation. It's a real product that can keep evolving as fast as the models it runs on.”
Let's talk
Get in touch
Contact us if you have any projects in mind
Tell us about your idea, expected budget, and timeline. We usually reply within one business day.