MUHAMMAD TAHA AMER

Summary

What I did

• Built AI voice agents for inbound appointment booking, lead qualification, and support using Vapi for orchestration.
• Configured n8n webhooks to trigger CRM updates, calendar bookings, and confirmations based on agent intent detection.
• Logged every call into Google Sheets and the CRM for monitoring and reliability tracking.
• Built an inbound-call automation in n8n where an LLM acted as a router to classify caller intent and extract data fields into a JSON schema.
• Designed multi-path routing for call outcomes including calendar bookings, CRM lead updates, Slack notifications, and support ticket creation.
• Implemented logic to propose alternative calendar slots via LLM if a requested time was unavailable.
• Developed a human-in-the-loop review queue for automated classifications with a confidence score below roughly 0.7.
• Integrated idempotency checks using call IDs to prevent duplicate bookings or lead entries during webhook retries.

Results

• Achieved an end-to-end response time under about 1 second for AI voice agents by using streaming at every stage and keeping prompts short.
• Moved slower tasks like CRM writes and calendar updates to asynchronous webhooks to prevent blocking real-time conversations.
• Implemented fallbacks, retries, and monitoring for failed automations to ensure reliability in voice agent pipelines.
• Targeted speech-to-text latency of 200-300 ms, LLM time-to-first-token of 300-400 ms, and text-to-speech of 150-250 ms.