Talk | ExpoStacker

LIVE PRODUCT

talk.expostacker.com.br

live preview

Talk — AI agents for WhatsApp

Use now Architecture

Tech stack: IA WhatsApp Business API Multi-tenant LLM RAG Cloudflare Workers

4
Plans: 6
Onboarding steps: 100%
Isolated tenant

Client

Talk

Sector

SaaS and Conversational AI

Timeline

Iteracao continua · 2026-09

Live product

Use now

Results

Client-reported results during rollout

4 Planos comerciais

6 Passos de onboarding

100 % Tenant isolado

Client-reported data

The challenge

Talk is ExpoStacker's platform for AI agents on WhatsApp Business. Unlike bots built on unofficial libraries — which end up banned by Meta — Talk runs on the official Cloud API: each company gets an isolated tenant, an agent with configurable tone and policies, its own knowledge base and a kill switch always at hand.

Pequenas e medias empresas querem atendimento automatico no WhatsApp, mas automacao com bibliotecas nao oficiais gera banimento da Meta. Era preciso um produto que usasse a API oficial, isolasse dados por empresa, desse controle real ao dono (kill switch, handoff, politicas) e respondesse com base no catalogo real — nunca inventando preco ou estoque.

The approach

Plataforma SaaS multi-tenant sobre a WhatsApp Business Cloud API oficial da Meta. Cada tenant tem agente de IA com guardrails: o LLM gera texto, mas codigo decide desconto, estoque e preco a partir do catalogo real (StackPost). Inbox unificada com estado AI/humano por thread, score de lead, handoff com contexto e resumo, knowledge/RAG por empresa, opt-out STOP/PARAR, supressao e kill switch por tenant. Onboarding self-service em 6 passos e planos comerciais via StackPost.

The solution

The product combines an LLM for natural language with deterministic code for business rules: the agent never invents price, discount or stock. A human takes over when needed — with context and summary ready.

AI agent with guardrails: LLM writes, code decides — discount, stock and price always from the real catalog
Official Meta API: 24h window, approved templates, STOP opt-out and suppression
Multi-tenant with full data isolation per company
Unified inbox with AI/human state per thread and lead scoring
Human handoff with context and automatic summary
Per-company knowledge/RAG: policies, FAQ and catalog
Per-tenant kill switch always within reach
Self-service onboarding in 6 steps, no technician needed
Commercial plans integrated with StackPost (BRL 197 to 1,497/mo + Agency)

Timeline

Como o projeto foi conduzido, do mapeamento ao deploy. How the project was conducted, from mapping to deploy.

Methodology

Como o projeto foi conduzido, do mapeamento ao deploy. How the project was conducted, from mapping to deploy.

Architecture

→ Next.js + Cloudflare Workers — frontend and API on the same edge runtime
→ Official Meta Cloud API for WhatsApp Business — no unofficial library
→ Multi-tenant Supabase with RLS — each client isolated per row
→ AI agent with RAG connected to the client's real catalog
→ Guardrails: configurable policies, kill switch and human handoff with summary
→ Commercial plans billed via StackPost/Billing in BRL

Conceptual diagram — does not reflect real implementation

Tech stack

Technical choices and what was left out.

Decision log

What I chose, what I rejected, and why.

Next.js 15 instead of a SPA

Why: I chose a hybrid rendering architecture for SEO, SSR and better initial performance — combinado com Next.js + Cloudflare Workers, Meta Cloud API oficial, Supabase (Postgres + Auth + RLS), AI + RAG + Guardrails.. A pure SPA would have hurt indexing and first-load experience.

Rejected alternatives: Traditional SPA with React Router — rejected due to lack of SSR and dynamic metadata.

PostgreSQL instead of MongoDB

Why: I picked a relational database because the catalogs, inventory and orders require consistency and well-defined relationships — combinado com Next.js + Cloudflare Workers, Meta Cloud API oficial, Supabase (Postgres + Auth + RLS), AI + RAG + Guardrails..

Rejected alternatives: MongoDB — rejected because the project data model benefits more from strict schemas and ACID transactions.

Tiered gamification instead of open leaderboards

Why: I built a closed tier system to recognize progress without exposing all users on public rankings — combinado com Next.js + Cloudflare Workers, Meta Cloud API oficial, Supabase (Postgres + Auth + RLS), AI + RAG + Guardrails., preventing early users from feeling demotivated.

Rejected alternatives: Open global leaderboard — rejected because it pressures beginners and hides individual progress.

Security and access

Lessons learned

Not everything went as planned.

— Sellers adopt when the manager adopts first

— Gamification works until it becomes routine — it needs seasonal mechanics

— Content curation is more important than the technology itself

Nexus IA

Nexus IA — Plataforma de IA

Plataforma de IA com 620+ endpoints REST, Nexus Auto com roteamento automático entre 20+ modelos, workspace com agente autônomo, multimodal e 6 planos comerciais.

View case →