No posts yet.
This is an app# Project Brief: AI-Powered Website Design & Build Studio ## What this is A local application that uses a team of AI agents to design, build, and refine complete websites automatically — starting from a business description and reference material, and ending with a working, deployed site. Think of it as an in-house "AI web design agency": instead of one AI generating a site in one shot, seven specialized AI agents each do a distinct job (planning, design, coding, testing, reviewing) in sequence, with real quality gates in between, so the output is closer to what a human dev/design team would produce than a single-prompt AI site generator. ## Who it's for The intended user is someone who wants to spin up a real, polished website for a client or their own business (e.g. a restaurant, a shop, a portfolio site) without hand-coding it — but wants more control and quality than a typical "AI website builder" SaaS product gives you. Two working examples already exist in the codebase: a redesign for an actual restaurant ("Green House on Nut") and a template rework for a business called "Tookata." ## How it currently works (technical shape) **Stack:** - Frontend/output: Next.js (React) site, styled and structured as a real e-commerce/business site template - Deployment target: Cloudflare Workers (with D1 database + Drizzle ORM for any dynamic data) - Orchestration: Python backend (`agent_team.py`) running the multi-agent pipeline - Local control panel: a Python-run web dashboard (`dashboard.py`, served at `127.0.0.1:8765`) where the user creates a new "client project," uploads research/reference images, and watches/reviews the pipeline's progress - AI providers: NVIDIA NIM (primary), with Google Gemini and Groq as fallbacks — no single vendor lock-in, so if one provider is degraded, the pipeline can fall back to another **The seven-agent pipeline:** 1. **Planner** — turns the business brief into a concrete site plan 2. **Concept Director / Designer** — explores visual directions; a "Concept Studio" mode generates multiple distinct visual concepts side-by-side so the user can pick a direction before full build-out 3. **Coder** — writes the actual site code, file by file (not one giant unreviewed dump) 4. **Tester** — runs a real `npm run build` against the generated code to catch broken builds before anything is shown to the user 5. **Repairer** — fixes build failures automatically 6. **Visual Reviewer** — takes real screenshots (desktop + mobile) of the built site and evaluates them 7. **Owner Reviewer / Final Reviewer** — acts as a deliberately picky "client," rejecting generic or low-quality output before final sign-off There's also a safety/guardrail layer that checks generated code against policy rules (blocked file types, size limits, forbidden paths like `.env`/`node_modules`) before it's ever applied to the project. ## Current state & known issues (relevant to anyone extending this) This is a working prototype, not a finished product. Known limitations currently being addressed: - **Attempt-budget architecture is too rigid.** The pipeline currently shares one small pool of "attempts" (3) across unrelated failure types — an expected owner-requested redesign, a guardrail rejection, and a genuine provider outage all draw from the same budget, so normal back-and-forth can exhaust the budget and kill a run that was actually mostly working. - **No versioning/rollback of in-progress builds.** The site is built into a single shared workspace directory across attempts, so a later failed attempt can overwrite an earlier, better one, with no way to recover it. - **Provider reliability.** NVIDIA's free/trial API tier has a track record (per their own developer forums) of intermittent 429/503/403 errors; this needs graceful fallback plus a way to verify provider health without having to manually test with `curl`. - **No provider health-check UI yet** — the plan is to add a "Test providers" button in the dashboard showing live connected/unauthorized/rate-limited/unavailable status per provider (NVIDIA, Gemini, Groq).
No posts yet.