customer-service

Customer Support Auto-Responder Agent

Production-grade AI-powered customer support with Agent OS governance.

🎯 Overview

This agent automatically handles customer support tickets with:

• Multi-model verification (CMVK) - Responses verified across GPT-4, Claude, Gemini
• RAG-based knowledge retrieval - 90% lookup, 10% reasoning
• Policy enforcement - No false promises, discount limits, professional tone
• Automatic escalation - Legal threats, complaints, sensitive topics
• CSAT tracking - Quality metrics and continuous improvement

Benchmark: "Resolved 2,400 tickets, 4.7/5 satisfaction"

🚀 Quick Start

# Install dependencies
pip install -r requirements.txt

# Run the demo
python main.py

📊 Features

Ticket Classification

• Billing, Technical, Refund, Account, Feature Request, Complaint
• Automatic priority assignment (Urgent, High, Medium, Low)

Policy Enforcement

Rule	Description
No Promises	Blocks "guarantee", "definitely", "100%"
Discount Limits	Tier-based: 10% (Tier1), 20% (Tier2), 50% (Supervisor)
Refund Limits	$100 (Tier1), $500 (Tier2), $5000 (Supervisor)
PII Protection	Redacts SSN, credit cards, passwords
Response Length	Max 500 characters

Escalation Triggers

• Legal threats: "lawsuit", "lawyer", "sue"
• Media: "twitter", "viral", "news"
• Executive: "CEO", "board"
• Sensitive: "injured", "discriminat"

Knowledge Base (RAG)

• Vector-based retrieval (Pinecone/ChromaDB ready)
• Pre-loaded articles for common issues
• Confidence scoring for response quality

📈 Metrics Tracked

tickets_handled: Total processed
tickets_resolved: Auto-resolved without human
tickets_escalated: Sent to humans
resolution_rate: % auto-resolved
avg_response_time_sec: Speed metric
avg_csat: Customer satisfaction (1-5)
policy_violation_count: Governance catches

🔧 Configuration

Edit SUPPORT_POLICY in main.py:

SUPPORT_POLICY = {
    "response_rules": {
        "no_promises": True,
        "max_response_length": 500,
    },
    "discount_limits": {
        "tier1": {"max_percent": 10},
    },
    # ...
}

🏗️ Architecture

┌─────────────────────────────────────────────────────────┐
│                    Customer Ticket                       │
└─────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                  TicketClassifier                        │
│         (Category + Priority Assignment)                 │
└─────────────────────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────┐
│                  PolicyEngine                            │
│            (Escalation Check)                            │
└─────────────────────────────────────────────────────────┘
                           │
              ┌────────────┴────────────┐
              ▼                         ▼
     ┌────────────────┐        ┌────────────────┐
     │  Escalate to   │        │  KnowledgeBase │
     │    Human       │        │   (RAG Search) │
     └────────────────┘        └────────────────┘
                                       │
                                       ▼
                            ┌────────────────────┐
                            │   PolicyEngine     │
                            │ (Response Check)   │
                            └────────────────────┘
                                       │
                                       ▼
                            ┌────────────────────┐
                            │  CrossModelVerifier│
                            │      (CMVK)        │
                            └────────────────────┘
                                       │
                                       ▼
                            ┌────────────────────┐
                            │   Final Response   │
                            │ (or Human Review)  │
                            └────────────────────┘

🔌 Integration

Zendesk

from zendesk import ZendeskAPI

zendesk = ZendeskAPI(api_key=os.getenv("ZENDESK_API_KEY"))
agent = CustomerSupportAgent()

for ticket in zendesk.get_new_tickets():
    response = await agent.handle_ticket(ticket)
    if not response.requires_human_review:
        zendesk.reply(ticket.id, response.content)

Slack

@slack_app.event("message")
async def handle_dm(event):
    ticket = Ticket(
        ticket_id=event["ts"],
        customer_id=event["user"],
        message=event["text"]
    )
    response = await agent.handle_ticket(ticket)
    slack_app.respond(event, response.content)

📋 License

MIT License - Use freely with attribution.