How we
use AI.

Who did what

The software architecture and the technology choices are mine, Narges Hadji-Hosseini. I am a software engineer with over 20 years of experience: rating models, investment banking, trading and ALM in the financial sector, plus Rust, Web3, smart contracts, cloud and AI features in production systems.

I decided how the backend is structured, which services are used, how data is stored, and where the line between the public and the protected area runs. Those decisions are the core, and they were not delegated.

The implementation was created together with two AI assistants: Claude and Codex. They wrote code, drafted copy, added tests and found bugs – always against my specification and reviewed by me.

How it is built

For the web server and the backend I deliberately chose Rust, with axum as the web framework. Not because it is fashionable, but because the compiler rules out an entire class of bugs before the code ever runs: no null pointers, no memory errors, no data races. After years on systems where a calculation error gets expensive, you learn to value guarantees like that. I am convinced by this choice, and it was mine.

The frontend is hand-written HTML and CSS with no build step – no site builder, no framework overhead. The public site is additionally delivered at the edge through Cloudflare so it stays fast everywhere; both implementations share the same routes and the same data model.

• Rust with axumWeb server and backend, built as a container – the deliberately chosen foundation
• SQLite / Cloudflare D1Leads, spaces and votes
• Cloudflare Pages & WorkersEdge delivery, close to visitors
• HTML & CSS without a frameworkNo build, no bundle, fast loads
• ResendTransactional email
• Claude APIDrafts for assessing member suggestions

Security was part of the architecture, not bolted on later: a strict content security policy, signed sessions, server-side validation of every input, and a protected area that reveals nothing without a valid session.

Where AI sits in the product

Nowhere at the moment. Exactly one place is planned: when a member suggests a space, a piece of furniture or a pricing model in the founding portal, Claude is to draft a first list of pros and cons from the description and, where present, the linked page.

That feature is switched off until the privacy notice covers the transfer to Anthropic. Once it runs, every draft will be visibly marked as an “AI draft” in the portal and never published automatically: every suggestion stays hidden until I have read it, corrected it and released it.

Where there is no AI

• Choosing the location
• Setting the prices
• The member vote – every vote comes from a person
• Deciding who is admitted
• Replies to personal enquiries

What it feels like

Honestly, working productively with AI still feels like a dream to me. I had long assumed it would come eventually – just not this fast. There is leverage that simply did not exist before: ideas can be tried out in hours instead of weeks, and the manual typing has all but disappeared.

Which is exactly why you have to stay the architect. The AI writes the code, but it does not decide how a system is built, what it has to do, or where its limits are. Give that role away and you end up with a lot of code and very little system. Keep it, and you work faster than ever.

Why we write this down

Because “built with AI” can mean anything today – from a carefully designed system to a copy-pasted page nobody checked. Anyone joining TechGarden should know which decisions a human made.

And because it is part of the concept: a workspace where AI belongs to the working day should be able to show what that looks like in practice – including the places where it has no business being.

Narges Hadji-Hosseini
Software Engineer & Founder