Paithra

paithra.com

Paithra is an autonomous AI voice calling agentic platform engineered by Pruning Labs. It handles multi-turn customer support and outbound qualification calls with sub-650ms turn-taking latency, telephony SIP trunking, and deterministic business logic guardrails on WebRTC, LiveKit, Cartesia, and Claude.

The problem

Call centers and operations teams face overwhelming call volumes, high agent turnover, and slow qualification speed. Rigid interactive voice response (IVR) phone trees frustrate customers and lead to high drop-offs.

Most generative voice bots suffer from high latency — two-second pauses feel unnatural — and lack deterministic guardrails, leading to unapproved commitments or hallucinated policy statements.

What we pruned

• Rigid multi-level IVR menus pruned in favor of fluid natural conversations
• Awkward 2-second audio latency pruned down to sub-650ms instant turn-taking
• Manual call tagging and post-call CRM logging pruned to zero

What we engineered

• Real-time streaming audio pipeline built on WebRTC and LiveKit
• Deterministic policy guardrails with typed action tools (calendar booking, CRM lookups)
• Bilingual and accented speech models tuned for Indian and global English speakers
• Automated call recording, structured summarization, and sentiment telemetry

How it is built

• Telephony audio connects via SIP trunking to a media gateway streaming audio bi-directionally over WebRTC.
• A turn-taking detector classifies conversational interruptions within 120ms to allow natural human interruptions.
• The reasoning engine executes on Claude 3.5 Sonnet with strict function-calling schemas for database updates and bookings.
• Audio synthesis streams ultra-low-latency voice chunks via Cartesia directly to the caller.

Stack

• WebRTC
• LiveKit
• Claude 3.5 Sonnet
• Cartesia
• FastAPI
• Twilio SIP
• PostgreSQL

Disciplines involved