Never Miss Another Lead Call

Overview

The CushLabs AI Voice Agent Platform is a production system for deploying conversational AI agents that handle inbound and outbound phone and web calls. Built on the Vapi real-time voice infrastructure, each agent combines an LLM brain (Claude Sonnet or Groq Llama 3.1), speech-to-text (Deepgram Nova-2/3), text-to-speech (Cartesia), and a webhook-driven backend that executes real-world actions — booking calendar appointments, writing lead data to a database, looking up property listings, querying business information — all during a live conversation.

Five demo agents are deployed at voice.cushlabs.ai: Clara handles lead qualification for CushLabs itself, James books executive coaching appointments with real-time Google Calendar integration, Sophia serves as a med spa front desk agent, Mike dispatches home service calls, and David — the platform's first outbound agent — proactively calls real estate prospects via Twilio PSTN to qualify buyers and book property tours.

The Challenge

• Lost leads: Unanswered calls during off-hours, lunch breaks, and peak volume mean revenue walks away
• IVR frustration: Menu trees and hold queues create friction that drives callers to competitors
• Staffing costs: A full-time receptionist costs $30-50K/year and still only covers business hours
• No intelligence layer: Traditional phone systems route calls but cannot qualify, score, or act on the conversation content

The Solution

Natural conversation, not menu trees: Each agent runs a full LLM with a domain-specific system prompt, handling free-form conversation rather than rigid scripted paths.

Real actions during calls: Vapi function-calling triggers webhook requests to the Express backend, which executes Google Calendar bookings, database writes, and business logic in real time — the caller hears confirmation within the same conversation.

Multi-agent architecture: Each demo agent has its own LLM, voice model, and system prompt optimized for its domain. Clara uses Claude Sonnet for nuanced qualification questions. James uses Groq for ultra-low-latency appointment booking. Sophia uses Claude with a detailed med spa knowledge base. Mike handles home service dispatch with scheduling logic. David makes outbound PSTN calls via Twilio to qualify real estate buyers.

Inbound + Outbound: Four agents handle inbound web calls via Vapi Web SDK. David introduces the platform's first outbound calling capability — a server-side endpoint triggers Vapi to call prospects via Twilio PSTN, with mock MLS property lookup and tour booking built into the conversation flow.

Session persistence: Redis stores conversation state across multiple function calls within a single session, allowing the agent to reference earlier context when executing later actions.

Technical Highlights

• Vapi WebRTC pipeline: Browser microphone → Deepgram Nova-2/3 STT → LLM (Claude/Groq) → Cartesia TTS → browser speaker, all under 500ms round-trip
• Outbound PSTN calling: Server-side POST /api/outbound-call triggers Vapi to call prospects via Twilio with E.164 validation and per-IP rate limiting
• Webhook function calling: Vapi triggers POST /api/webhook with structured function call payloads; server routes to calendar, database, MLS lookup, or custom business logic services (8 function handlers)
• Mock MLS integration: Property lookup by ID or address for real estate demo with 6 NJ listings
• Google Calendar OAuth: Real-time availability checks and event creation with auto-generated Google Meet links during live calls
• Redis session state: Maintains conversation context across multiple function calls within a single Vapi session
• Neon PostgreSQL persistence: Lead data, call metadata, and booking confirmations stored for CRM integration
• Full bilingual i18n: Client-side EN/ES toggle with localStorage persistence, MutationObserver for dynamic Vapi status translation
• Render deployment: Express server + Redis instance via Render Blueprint (render.yaml) for reproducible infrastructure

Results

For the Business:

• Every inbound call answered instantly, 24/7 — zero missed leads
• Qualified prospects booked directly into calendar before hanging up
• L1 support handled automatically, human agents reserved for complex issues

Technical Demonstration:

• Real-time voice AI with production-grade latency (<500ms)
• Both inbound (Web SDK) and outbound (Twilio PSTN) calling in a single platform
• Webhook-driven architecture that executes real-world actions during live calls (8 function handlers)
• 5-agent deployment across 5 industries with domain-specific LLM, voice, and system prompt configurations
• Full bilingual EN/ES support with client-side i18n
• Infrastructure-as-code via Render Blueprint for reproducible deployments