Phonx AI – Real-Time Voice AI Platform
Problem
Insurance agencies were overwhelmed with manual calls to answer complex benefit questions. Existing systems offered limited automation and could not maintain regulatory compliance in long-form conversations, leading to bottlenecks and inconsistent customer experiences.
Context / Business Need
Phonx AI set out to build a voice automation system that would reduce human call load, ensure compliance with insurance regulations and deliver responses in under one second. The platform needed to handle long-form interactions, preserve context across multiple turns and integrate seamlessly with CRM and telephony infrastructure.
Constraints
- Regulatory constraints requiring compliance and auditability of all conversations.
- Sub-second response times for real-time voice interactions.
- Accurate long-form conversation handling across multiple call turns.
- Integration with legacy telephony and CRM systems.
- Strict data privacy and security requirements in the insurance domain.
My Role
As AI Product Engineer and Head of Development, I led requirements elicitation, designed the system architecture and oversaw the development team. I ensured regulatory compliance by embedding a domain-specific knowledge layer and collaborated closely with legal and insurance experts.
System-Thinking Approach
We mapped the user journey from the moment a policyholder called to final resolution. Each component — speech-to-text, conversation manager, knowledge retrieval, compliance layer, CRM connector and telephony gateway — was designed to support one another, enabling fluid, compliant conversations.
MVP Design
The MVP focused on core elements: high-accuracy STT/TTS pipelines, a proprietary knowledge engine for insurance content, a compliance module, CRM integration and conversation memory to track context. We avoided feature creep to ensure speed and reliability.
Architecture Breakdown
The system comprised:
- A telephony gateway handling SIP calls and routing.
- STT and TTS microservices converting audio to text and back.
- A conversation manager orchestrating LLM reasoning and maintaining state across long calls.
- A compliance module enforcing insurance regulations and logging decisions.
- A CRM connector syncing policy data and recording interactions.
- A proprietary knowledge retrieval engine providing accurate answers in real time.
- A microservice event bus enabling sub-second response times.
Final Solution & Results
The final platform achieved average response times under one second and 90%+ intent accuracy. It reduced human call volume by over 60%, provided consistent, compliant answers and allowed insurance agencies to scale support without growing staff.
Tech Stack
- Python, Go and Node.js microservices
- LLM orchestration with custom insurance knowledge engine
- Docker & Kubernetes for scalability
- gRPC, event buses and real-time telephony APIs
- Twilio, proprietary STT/TTS engines