Case Study 04 · AI Automation
AI Voice Receptionist
End-to-end AI automation for inbound business calls, qualification, and CRM logging
Intro
Product type: Productised AI system, live on this portfolio as Liya.
Use case: Inbound call handling for businesses that need 24/7 lead qualification without a receptionist.
Role: Sole architect and engineer, voice agent design, N8N automation pipeline, webhook routing, and CRM integration.
What Made This a Good Bet
Businesses running inbound campaigns miss or mishandle calls outside business hours, losing high-intent leads.
Hiring and training human receptionists is expensive and introduces inconsistency in qualification conversations.
Existing IVR and chatbot tools were scripted and non-conversational, causing leads to drop off early.
The breakthrough isn't the model alone, it's the operational pipeline behind it. Voice without logging, retries, and deduplication is only a prototype.
Every missed call is a missed opportunity; for high-ticket offers, one missed call can represent thousands in lost revenue.
Manual CRM entry by staff introduced errors and delays in critical follow-up tasks.
Businesses needed a system that was always on, always consistent, and automatically logged structured data.
What I Built
Built a full AI voice agent using Vapi with a natural conversation flow for lead qualification and appointment scheduling.
Designed an N8N automation pipeline that handles booking tool calls, structured data extraction, and real-time CRM logging.
Engineered a webhook router that filters and routes 6+ Vapi event types with zero duplicate processing.
Integrated with Airtable as the CRM backend, with per-call transcripts, lead scores, and contact records.
The same system powers the AI assistant visible on this portfolio.
The Stack
The goal was validation speed first, then reliability: get a real assistant on a real phone number, observe failure modes in production telemetry, then harden the automation layer once traffic proved the workflow.
Why Reliability Matters More Than a Flashy Demo
A conversational voice layer is worthless if duplicates hit the CRM, if booking tools race, or if transcripts arrive out of order. Those edge cases dominate real-world behavior once you leave scripted demos.
This build treats event ordering, idempotency, and structured payloads as first-class citizens, the same mindset that scales from one demo line into production traffic.
That operational discipline is what makes the assistant viable for business use, moving beyond simple landing page headlines.
