Ask a customer to describe their business in a sentence and, a few seconds later, hand them a working website — that was the product bet. Delivering it meant turning an LLM into a reliable site-generation engine that stays responsive while it works, produces consistent structure every time, and holds up when hundreds of people generate at once.

I led the Name.am AI website builder end-to-end as the lead full-stack and AI engineer — from the generation architecture to the streaming UI. The builder takes a short prompt and produces a structured, editable website, generating content and layout through a large language model rather than fixed templates.

• DeepSeek LLM integration as the generation engine, chosen for cost-to-quality at scale
• Streaming output over Server-Sent Events (SSE) so users watch the site build in real time instead of staring at a spinner
• A modular prompt system that composes generation from reusable, individually testable prompt units
• Concurrency-safe architecture for many simultaneous generations
• 1,000+ sites generated and a ~40% reduction in build time over the baseline flow

The core UX decision was to stream. Full-site generation takes real time, and a blank loading state at that length reads as broken. Streaming results over Server-Sent Events turns the wait into progress: the user sees sections appear as the model produces them, and perceived speed improves even when total generation time is unchanged.

Behind the stream, generation is composed from a modular prompt system rather than one monolithic prompt. Each unit is responsible for a slice of the output and can be tuned and tested on its own, which keeps generated structure consistent and makes the system safe to iterate on — changing how one section is generated doesn't destabilize the rest.

To hold up under real traffic, the generation pipeline is built to run many jobs concurrently without their streams interfering, so throughput scales with demand instead of serializing behind a single request.

Prompt-to-site generation — A short business description produces a structured, editable site: sections, copy, and layout, generated rather than picked from a fixed template.

Real-time streaming preview — SSE streams the model's output to the browser as it's produced, so generation is visible and responsive from the first token.

Modular, testable prompts — Composing generation from small prompt units keeps output consistent and makes the system iterable without regressions.

Built for scale — Concurrency-safe streaming supports many simultaneous generations, which is what made 1,000+ generated sites practical.

• Streaming is a product feature, not a technical detail. The same generation time feels dramatically faster when the user watches it happen. Server-Sent Events were the cheapest, most reliable way to make a slow-by-nature operation feel alive.
• Modular prompts are how you keep an LLM predictable. One giant prompt is impossible to reason about; decomposing generation into small, testable units is what kept output consistent and let me improve one part without breaking the others.
• Concurrency is the difference between a demo and a product. Making generations run in parallel without their streams colliding is what turned a neat prototype into something 1,000+ real sites could be built on.