💬
Live in productionProductVery high complexity

ChatDemos AI

Universal WhatsApp Business AI platform with strict Human-in-the-Loop enforcement.

0.5–1 s
Response time
60–75 %
Cost reduction (vs n8n)
1,000+
Concurrent users
$5–7 / mo
Operational cost

What is ChatDemos AI?

ChatDemos AI is a production-ready, single-tenant WhatsApp AI conversation platform with strict Human-in-the-Loop enforcement. It is designed to be deployed as separate isolated instances per business — each instance owns its own data, knowledge base, conversation history, and HITL approval flow.

The core architecture is WhatsApp → Pydantic AI (FastAPI) → Supabase → Dashboard. AI replies are generated by Claude 3.5 Haiku via OpenRouter, cached semantically (60–73% cost reduction), and either auto-sent or held for human approval depending on per-business policy.

It is the single largest production system shipped end-to-end out of ClosedChats AI to date — and it serves as the architectural reference for every WhatsApp + AI engagement we ship.

Who is it for?

  • SMBs that handle most customer conversations on WhatsApp
  • Verticals where AI cannot be unsupervised: healthcare, real estate, regulated finance
  • Operators who need true data isolation per business (not multi-tenant SaaS)
  • Teams that want production telemetry from day one — not a black-box chatbot

What does it do?

  • WhatsApp Business API ingest + reply
  • Real-time HITL approval dashboard (human reviews before send)
  • Per-business knowledge base (products, FAQs, documents) with vector search
  • Semantic response cache (60–73% LLM cost reduction)
  • Jailbreak / spam / rate-limit security layers
  • Real-time analytics on conversation volume, intent, deflection rate
  • Single-tenant deployment template (one Coolify project per business)

How is it built?

PythonFastAPIPydantic AIClaude 3.5 HaikuOpenRouterSupabase (Postgres)Next.js 16TypeScriptCoolifyTraefikCloudflareAxiom (telemetry)

What makes it interesting?

  • 60–75% latency improvement over the equivalent n8n workflow
  • HITL is enforced architecturally — not bolted on as a setting
  • Semantic cache cuts LLM cost by 60%+ without hurting answer quality
  • Single-tenant by design: each customer's data is in its own database
  • Production-tested at 1,000+ concurrent users with full observability

Live workflows we ship like this

Production automation patterns we deploy for clients on the same primitive — customized to their stack on the engagement.

Alex — WhatsApp Sales Agent — workflow diagram
Sales

Alex — WhatsApp Sales Agent

Intelligent sales agent that routes, prioritizes, and resolves customer inquiries. AI-powered product suggestions, sentiment analysis, in-conversation payment capture, and CRM hand-off — all through one WhatsApp number.

Integrations
WhatsAppPostgreSQLChatGPTOpenAI

Want something like this for your business?

We adapt the patterns above to your stack on every engagement. The 15-minute discovery call is free — you leave with a plan regardless.