High-Level Design

How AGX Chat is built.

A multi-tenant conversational AI platform: one brain across every channel, three engines (SELL · SUPPORT · SCALE), grounded knowledge, secure payments and enterprise integrations — configured per customer with a single config file.

Three engines, one platform

01 · SELL

Revenue Engine

Right product at the right moment · value-based selling · urgency offers · in-chat payment · automated follow-up.

02 · SUPPORT

Resolution Engine

Resolves 80%+ unaided · RAG answers with citations · sentiment detection · context-rich human handoff · 24/7 multilingual.

03 · SCALE

Growth Engine

Any CRM/ERP/commerce · 10+ channels at once · 1→100k conversations · multi-tenant · live in weeks.

Container architecture

Channels normalise into one gateway; the AutoGenX orchestrator drives skills, knowledge, sentiment, offers and payments; everything syncs to data stores and external systems.

ChannelsTouchpoints
🟢 WhatsApp
💬 Web widget
📸 Instagram
📘 Messenger
📞 Voice
💬 SMS
✉️ Email
▼
GatewayNormalise & identify
Channel adapters
Webhook auth · dedupe
Unified message envelope
Customer identity
▼
OrchestratorAutoGenX core "brain"
Intent + language
Dialog state machine
Policy & guardrails
Tool / function calling
LLM gateway (+fallback)
▼
Skills & servicesCapabilities
SELL playbook
SUPPORT playbook
SCALE playbook
RAG engine (cite)
Sentiment & handoff
Offer / pricing
Payment sessions
▼
Data & integrationState & systems
Conversation store
Customer profile (CDP)
Vector store (RAG)
Analytics stream
CRM · ERP · Commerce
Payments
▼
Tenant surfacesFor your team
Agent console
No-code studio
Automations
Analytics dashboards
Multi-tenant by design

One config = one customer.

Every deployment — Travel, Real Estate, Retail, or a real client — is a single declarative config: brand, catalogue, knowledge base, offer and an AI persona that scopes the assistant. Onboarding a new customer is adding one file.

  • Brand & theme (colours, logo, voice)
  • Catalogue (products / listings / tours)
  • Knowledge base (RAG sources + citations)
  • Offers, suggestions & channels
  • AI persona & scope guardrails
// server/scenarios/acme.js
export default {
  id: "acme",
  label: "ACME Co.",
  brand: { name, logo, colors, assistant },
  catalogue: [ /* products */ ],
  kb:        [ /* RAG Q&A + sources */ ],
  offer:     { code, label, detail },
  persona:   {
    business: "...",
    goals:    "sell, support, capture leads",
    style:    "on-brand, concise"
  }
}

Scoped, grounded AI

The chat is powered by an LLM (OpenAI) behind a server-side proxy — the API key never reaches the browser. Each request is constrained to its scenario.

🎯 In scope only

A per-scenario system prompt limits the assistant to that business, its catalogue and knowledge base. Off-topic questions are politely redirected.

🧱 Grounded

It can only recommend real catalogue items (by id) and answers factual questions from the knowledge base — no invented products, prices or policies.

🔧 Structured actions

The model returns JSON the UI renders: which product cards to show, whether to present the offer, capture a lead, or hand off to a human.

No API key configured? The widget automatically falls back to a scripted engine, so demos never break.

Key flows

SELL

Conversation → paid order

Message arrives on any channel; customer identified.
Orchestrator detects intent + language.
Catalogue tool returns matches → value-led cards.
On hesitation, Offer Engine issues a time-boxed discount.
Secure in-chat payment / booking created.
Order → CRM, confirmation cross-channel, follow-up scheduled.
SUPPORT

Question → cited answer → safe handoff

Question intent → RAG retrieves top-k chunks.
Answer grounded in sources; citation shown.
Sentiment scored every turn.
Negative threshold → escalation raised.
Human gets full transcript + profile + reference.
Retention offer optionally applied.

Deployment

Node.js + Express app, containerised, behind nginx with HTTPS — running at demo.autogenx.ai on an AWS Sydney (ap-southeast-2) EC2 instance.

🟢

Node + Express

Serves the UI + the scoped /api/chat proxy.

📦

Docker

Reproducible image; one command to run.

🔒

nginx + TLS

Reverse proxy with Let's Encrypt HTTPS.

☁️

EC2 · Sydney

ap-southeast-2, GoDaddy DNS A record.

Full runbook: DEPLOY.md · written design: docs/AGX_Chat_HLD.md

← Back to the live demo All products