AI Calling Agent
Bidirectional voice AI over real telephony for real estate
An AI calling platform that handles both directions of a real-estate business's phone line: bulk outbound campaigns and inbound customer calls run through a single LiveKit SIP agent that detects call direction and switches personas accordingly. Every call is transcribed, sentiment-analyzed and scored Hot/Warm/Cold into a live dashboard.
The Problem
A real-estate firm needed sales outreach at a scale human teams can't sustain (plus 24/7 inbound handling) over actual phone numbers in India, where SIP provider choice (Airtel IQ, Exotel vs Twilio) materially changes cost and reliability.
The Solution
One agent, two personas: room metadata tells the agent whether it's dialing out (campaign scripts: sales, reminders, surveys) or answering in (support persona). The voice pipeline pairs Deepgram STT and Silero VAD with GPT-4 and OpenAI TTS. Post-call, transcripts land in Firebase where AI post-processing produces sentiment and lead scores; a FastAPI webhook server and monitoring dashboard complete the loop.
Architecture
- LiveKit Agents + SIP trunking, provider-abstracted across Twilio, Plivo, Airtel IQ and Exotel
- Direction detection from room metadata → persona and prompt selection
- Deepgram STT + Silero VAD + GPT-4 + OpenAI TTS pipeline
- Firebase transcript storage → AI sentiment analysis → Hot/Warm/Cold lead scoring
- Dockerized agent and webhook services with rate-limited campaign scheduling
Challenges & How I Solved Them
Telephony in India is not Twilio-shaped
Indian numbers need TRAI-compliant providers; the SIP layer was abstracted so Airtel IQ/Exotel and Twilio are interchangeable, documented with a cost comparison per provider.
One agent, opposite jobs
Outbound selling and inbound support need different tone, compliance and interruption behavior, solved with direction-aware prompt selection rather than two deployments.
My Contributions
- Entire platform: agent, SIP setup, campaign scheduler, scoring pipeline, dashboard
Impact
- Bulk outbound campaigns and 24/7 inbound support on one codebase
- Automatic lead qualification (Hot/Warm/Cold) from call sentiment
Lessons Learned
Real-world voice agents live or die on telephony plumbing: VAD tuning, provider failover and direction-aware behavior matter more than the LLM.