Description

This project implements a multi-agent customer support system for a fictional mobile carrier called FM.inc, integrated with WhatsApp via Kapso. The system routes incoming messages through a LangGraph StateGraph, dispatching them to specialized agents: one for product/service information (RAG + Pinecone) and one for technical support (tool-based ticket management via PostgreSQL). Every conversation is persisted per user phone number using an AsyncPostgresSaver checkpointer.

A complementary n8n workflow extends the system by listening to real-time database changes and notifying the support team via email, enabling them to manage ticket status directly from their inbox.

graph TD
    WA["📱 WhatsApp"] --> K["Kapso"]
    K -->|POST /webhook| API["FastAPI"]
    API -->|BackgroundTask| G["LangGraph StateGraph"]

    G --> T["🔀 Triage"]
    T -->|"intent: info"| IA["📚 Info Agent<br/>(RAG + Pinecone)"]
    T -->|"intent: soporte"| SA["🔧 Support Agent"]

    SA -->|tool_calls?| TC{"tools_condition"}
    TC -->|sí| TOOLS["⚙️ ToolNode<br/>(inject phone)"]
    TOOLS -->|resultado| SA
    TC -->|no| FR["📝 Format Review"]

    IA --> FR
    FR -->|respuesta| SUM{"should_summarize?"}
    SUM -->|">6 msgs"| SC["🧠 Summarize"]
    SUM -->|"≤6 msgs"| REPLY["📱 WhatsApp Reply"]
    SC --> REPLY

    style T fill:#f59e0b,color:#000
    style IA fill:#3b82f6,color:#fff
    style SA fill:#8b5cf6,color:#fff
    style TOOLS fill:#6366f1,color:#fff
    style FR fill:#10b981,color:#fff
    style SC fill:#ec4899,color:#fff

Tech Stack

• Orchestration: LangGraph
• LLM: OpenRouter → OpenAI GPT-OSS:120B
• Vector DB: Pinecone Serverless
• Embeddings: Pinecone Inference (llama-text-embed-v2, 1024 dims)
• Relational DB: PostgreSQL (Docker)
• Memory: LangGraph AsyncPostgresSaver (per thread_id / phone number)
• API: FastAPI + Uvicorn
• WhatsApp: Kapso (WhatsApp Cloud API proxy)
• Automation: n8n
• Observability: LangSmith, Adminer
• AI-Assisted Development: Google Antigravity (scaffolding leveraging AI Skills and debugging optimization).

How It Works

Demo

1. Message Intake

Every WhatsApp message arrives as a POST request to the FastAPI /webhook endpoint via Kapso. The user's phone number is extracted and stored in the shared state — it will be used later to identify them in the database without ever trusting the LLM with it.