AI Receptionist & Lead Automation

Most small businesses miss 30-50% of inbound calls during operating hours. When the owner is on a job site, in a meeting, or asleep — the phone rings to voicemail. The lead goes cold, the customer calls a competitor, and revenue is lost.

Existing solutions were either expensive enterprise call centres, rigid IVR trees that frustrate callers, or simple voicemail. I wanted a system that could hold a natural conversation, extract structured information, and integrate with existing business workflows — all self-hosted and controllable.

This project explores: Can a local LLM + voice interface + workflow automation handle real business conversations reliably?

The architecture follows a clean event-driven pattern:

┌─────────────┐     ┌──────────────┐     ┌─────────────┐
│   Caller    │────▶│   Dograh     │────▶│    n8n      │
│  (Phone)    │     │  (Voice GW)  │     │ (Orchestrator)│
└─────────────┘     └──────────────┘     └──────┬──────┘
                                                 │
                    ┌──────────────┐     ┌───────▼───────┐
                    │  Webhooks    │◀───▶│  Local LLM    │
                    │  (Calendar,  │     │  (llama.cpp)  │
                    │   CRM, etc)  │     └───────────────┘
                    └──────────────┘            │
                                                ▼
                    ┌──────────────┐     ┌─────────────┐
                    │  Notifications│◀───▶│   Baserow   │
                    │ (Slack/Email) │     │  (Leads DB) │
                    └──────────────┘     └─────────────┘

Data Flow

1. Inbound call hits Dograh SIP trunk
2. Dograh performs STT → streams text to n8n webhook
3. n8n passes conversation context to local LLM
4. LLM responds with structured action: {intent, entities, response_text, tool_calls}
5. n8n executes tool calls (calendar, Baserow, notifications)
6. Response text → TTS → back to caller via Dograh
7. Full transcript + extracted data saved to Baserow

This is a working prototype — not a production deployment with client metrics. What it proves:

• Local LLMs (7B-13B) can handle structured conversation extraction reliably with good prompting
• n8n + Dograh + llama.cpp is a viable self-hosted voice AI stack
• End-to-end latency < 3s (STT → LLM → TTS) is achievable on modest GPU
• Business logic in n8n enables rapid iteration without code deployments

Next Steps

• Add RAG for business-specific knowledge (pricing, policies, technical specs)
• Implement caller identification via CLID + CRM lookup
• Add multi-language support (TTS/STT models for target languages)
• Build admin dashboard for non-technical script management
• Load test with concurrent calls to find bottlenecks