Sevri — MVP

A real, working version of the Sevri flow: sign up, tell the AI about your business in a chat, get a live page, and manage it (plus a simulated marketing chat) from a dashboard.

This is an MVP of the core path. AI replies, campaign drafts, and ad performance are simulated — see "What's simulated vs. real" below.

Stack

• Next.js 14 (App Router) + TypeScript + Tailwind CSS
• Database: Postgres, hosted on Neon, connected via Vercel's Storage integration. Accessed through @neondatabase/serverless (HTTP driver, works from serverless functions). Schema is created automatically on first query — no manual migration step.
• Auth: custom email/password auth. Passwords are hashed with scrypt (Node's crypto module); sessions are a signed, HTTP-only cookie (HMAC-SHA256, no third-party auth library).
• No external API calls anywhere in the app — it runs fully offline once installed, aside from loading Google Fonts in the browser (cosmetic; the app works fine without them).

Running it

npm install
npm run dev     # http://localhost:3000

For a production build:

npm run build
npm start

For local development you need a DATABASE_URL pointing at the Neon database. Run vercel env pull .env.local (after vercel link) to fetch the real one from this project, or paste a Neon connection string into .env yourself. The first request creates all tables automatically (CREATE TABLE IF NOT EXISTS).

Environment variables (.env)

• DATABASE_URL — Postgres connection string (Neon). In production this is injected automatically by Vercel's Neon Storage integration; for local dev, pull it with vercel env pull .env.local or set it by hand.
• SESSION_SECRET — HMAC signing key for session cookies. Change this before deploying anywhere real — the checked-in value is a dev-only placeholder.

The flow

1. /signup — create an account (email + password).

2. /start — "How do you want to start?" Pick one of two paths (this is the real, wired-up version of the prototype's Bifurcación screen), then press Continue:

   • "I already offer a service" → /onboarding — a short guided chat: business name, category, city, and a one-line description.
   • "I want to offer one, but I'm not sure what" → /discover — a longer, multi-screen path, all real (no hardcoded example business):
     1. A discovery chat (location, background/experience, licenses, time & tools).
     2. An analysis screen — "local demand," "your fit," and "startup cost" cards generated from what you actually typed, plus a "what we ruled out" section that filters out any suggested idea that would need a license you said you don't have.
     3. Three suggested service ideas, picked from a small rule-based idea bank keyed off your background text (construction/handyman, cleaning, landscaping, or a general fallback) — with a "show me other ideas" alternate set.
     4. A second chat to nail down the business name, pricing, service area/contact preference, and whether you have real project photos.
     5. A proposal screen summarizing all of it, with a field to add your real WhatsApp number, before it creates your business for real.

   Either path ends by writing a real Business row (pitch, starter service, unique slug) — not a mock — and lands on /dashboard.

3. /dashboard — two tabs:

   • Website: your generated page's content, a link to the live page, and a field to save your WhatsApp number (the live page's contact button uses it).
   • Marketing: a chat where you describe a goal ("get more weekend bookings") and Sevri drafts a campaign (ad copy + a suggested budget) and saves it as a draft under "Your campaigns."

4. /site/[slug] — the public generated page for the business, with a WhatsApp deep link if a number was saved.

What's simulated vs. real

This was intentionally scoped as an MVP of the product flow, with real integrations left as a clearly-marked next step rather than faked:

• Sevri AI's replies (lib/ai.ts) are template-based, not calls to a real language model. Every function there is small and pure on purpose — swap the body of generatePitch, nextOnboardingPrompt, nextDiscoveryPrompt, generateAnalysis, generateIdeas, and draftCampaign for real LLM calls (OpenAI, Anthropic, etc.) without touching any UI or database code. The discovery path's idea bank (IDEA_BANK in lib/ai.ts) is a small, rule-based set of options per skill cluster — a real model would generate these live instead of picking from a fixed list.
• Campaigns are always created as drafts. Nothing calls the real Meta Ads API. The chat reply says this explicitly rather than pretending a campaign is live.
• Reviews are never auto-generated. The database has a reviews table and the public page renders them, but nothing writes fake reviews into it — a real business page should never show testimonials that didn't happen. Wire up a real review-collection flow before this table is used for anything user-facing.
• No WhatsApp Business API, Meta Ads, or Google Business Profile integration — the "Message on WhatsApp" button is a real wa.me deep link (that part works today with no API key), but there's no server-side integration with any of those platforms yet.

Known limitations before a real deployment

• npm audit will flag Next.js 14.2.x against several advisories that are only fully patched in Next.js 15/16. Most don't apply to this app's usage (no custom server, no Image Optimization remote patterns, no i18n middleware), but upgrading before a public production deploy is recommended.
• Google Fonts are loaded via a <link> tag (like the original design prototype) rather than next/font, so they need outbound internet access in whatever environment serves the app — the app itself has no other network dependency.

Project layout

app/
  actions.ts            server actions: signup, login, both onboarding paths, WhatsApp, marketing chat
  page.tsx               marketing home page for the app itself
  signup/, login/        auth forms
  start/                  "How do you want to start?" — picks between the two paths below
  onboarding/             path A: "I already offer a service" chat + page wrapper
  discover/               path B: "not sure what" — discovery chat → analysis → ideas → detail chat → proposal
  dashboard/              Website / Marketing tabs, marketing chat
  site/[slug]/            the public generated business page
lib/
  db.ts                  Postgres (Neon) schema + all data access functions
  auth.ts                password hashing + session token sign/verify
  session.ts             reads the current logged-in user (+ their business) from the cookie
  ai.ts                  all "Sevri AI" logic, both paths — swap for a real model here