An agentic AI concierge — and three connected marketplaces — for stays, dining, rides & travel across Morocco.

 

---

OMNIA inverts the usual "super-app" of forms. You don't search — you make a wish. Describe what you want in plain language — "plan my weekend in Marrakech for two: a stay, dinner and an airport ride, under 3,000 MAD" — and the agent decomposes it, compares real options across three genuinely separate marketplaces, respects your budget, and assembles a confirm-ready plan.

Designed, built, deployed & documented end-to-end (product, frontend, backend, AI, infra) by El Asmi Ilyas Daoud.

🌐 Live

Surface	URL
🪄 OMNIA Agent	https://omnia-vert.vercel.app
🏠 OMNIA Stays	https://omnia-stays.vercel.app
🍽️ OMNIA Eats	https://omnia-eats.vercel.app
🚗 OMNIA Rides	https://omnia-rides.vercel.app

📸 A look inside

The cinematic “eye-portal” hero — state a wish.	Restaurant suggestions, with the agent’s live plan & execution log.
The in-chat menu cart — add several dishes, then review one order.	OMNIA Stays — handpicked riads, lofts & villas.
OMNIA Eats — full menus, live delivery info.	OMNIA Rides — six tiers, live distance-based fares.

🧩 The four apps

App	What it is	Theme	Frontend	Backend
🪄 OMNIA Agent	AI concierge — chat + live agent activity	Gold / black	ai-agent-hub/ (5180)	ai-agent-hub/backend/ (3000)
🏠 OMNIA Stays	Airbnb-style stays marketplace	Blue	omnia-stays/frontend/ (5181)	omnia-stays/backend/ (3001)
🍽️ OMNIA Eats	Food-delivery marketplace	Orange	omnia-eats/frontend/ (5182)	omnia-eats/backend/ (3002)
🚗 OMNIA Rides	Ride-hailing marketplace	Green	omnia-rides/frontend/ (5183)	omnia-rides/backend/ (3003)

Plus omnia-e2e/ — a Playwright end-to-end suite covering all four apps.

It serves six Moroccan cities: Rabat, Casablanca, Marrakech, Tanger, Oujda, Agadir. Currency is MAD.

🤖 What the agent can do

• Natural-language ordering & booking — "order a Margherita and a Coca-Cola from Napoli in Rabat" → a confirm-ready receipt that creates a real order (tagged via OMNIA Agent).
• Full-trip planning across all three marketplaces — stay + ride + meals assembled into one itinerary, with a hard budget cap.
• A live menu cart — suggested restaurants open their real menu in-chat; add several dishes with quantity steppers, then review one combined order.
• Manage existing orders — cancel, modify nights/guests, change a ride's pickup, edit order items, reorder "your usual".
• Recurring tasks — "order my usual from Dar Tagine every Friday at 7pm"; a cron fires it while you're away.
• Memory & context — remembers the conversation; follow-ups like "just a water please" keep the active restaurant.
• Scope guardrails — politely declines off-topic requests and unsupported cities.
• Reviews, an OMNIA Rewards wallet, live ride tracking, maps, and a per-user request meter.

🏗️ Architecture

        ┌──────────────────────── shared Clerk identity ────────────────────────┐
        │                                                                        │
  OMNIA Agent ──HTTP──▶ OMNIA Stays / Eats / Rides marketplaces (each its own DB)
  (chat + planner)        book / order / ride  →  real rows, "via OMNIA Agent" badge
        │
   Groq (Llama 3.3 70B) with multi-key failover

• Frontend: React 19 + Vite + TypeScript + TailwindCSS + Framer Motion + Zustand + React Query.
• Backend: NestJS + Prisma + PostgreSQL (Neon), one database per app.
• AI: Groq (Llama 3.3 70B) via a provider-agnostic layer, with automatic multi-key rotation and transient-failure retries.
• Auth: Clerk (shared across all four apps) with Google sign-in.
• Realtime: the agent streams its plan, steps, tokens and checkout drafts over Server-Sent Events.
• Infra: front-ends on Vercel, the four Dockerised backends on Hugging Face Spaces, databases on Neon — auto-deployed from this repo via GitHub Actions. The whole production stack runs on a $0 budget.

The agent is a hybrid: the LLM plans and phrases, but a large deterministic engine resolves every real order — so it never hallucinates a price, a place, or a booking.

See OMNIA_OVERVIEW.md for the full architecture, run guide, and security checklist.

⚙️ Run it locally

Each app needs its own .env (copy from the .env.example in each backend/frontend and fill in your own keys — Neon database URLs, Clerk keys, and a Groq API key).

# 1. Install deps (per app)
npm --prefix ai-agent-hub install && npm --prefix ai-agent-hub/backend install
#   …repeat for omnia-stays, omnia-eats, omnia-rides (frontend + backend)

# 2. Push the Prisma schema + seed each backend's database
npm --prefix ai-agent-hub/backend run prisma:push

# 3. Build + start the backends under PM2
npm --prefix ai-agent-hub/backend run build   # repeat per backend
pm2 start ecosystem.config.cjs

# 4. Start the frontends (Vite dev), ports 5180–5183
npm --prefix ai-agent-hub run dev

🔐 Security

• All real secrets live in per-app .env files, which are gitignored — only .env.example templates are committed.
• Per-user (or per-IP) rate limiting on the agent endpoint: 20/min + a daily cap persisted in Postgres (survives free-host restarts).
• Helmet headers, exact CORS allow-lists, per-user request keying by Clerk sub, and strict input validation at every API boundary.

👤 Creator

Built with ♥ in Morocco by El Asmi Ilyas Daoud — full-stack engineer · AI agents · product design.