CLARA: a unified hybrid ai-human receptionist framework using WebRTC and Monorepo architecture

Abstract

The rapid digitalization of enterprise infrastructure has caused a major change in institutional communication. This change requires the traditional front desk to transform into a high-tech center for security and logistics. However, current automated solutions often create what is known as the “automation paradox.” In this situation, removing human oversight increases the importance and complexity of the remaining interactions. This work outlines the design, implementation, and evaluation of conversational low-latency AI receptionist agent (CLARA). CLARA is a new hybrid system which solves the problems of unattended kiosks by combining a conversational AI, which uses the Google Gemini API, with a low-latency, real-time WebRTC video calling system. The design features a strong 3-Tier Monorepo built with React, TypeScript, Express.js, Socket.IO, and PostgreSQL. This setup ensures type safety and consistency in the code across the entire stack. Key contributions of this project include a custom Socket.IO signaling server for secure peer-to-peer connections, a “Human-in-the-Loop” workflow that reduces handoff issues, and an “Offline-First” strategy ensuring the system keeps working during network disruptions. Performance tests show that CLARA achieves less than 200 ms glass-to-glass latency and handles high traffic well, proving it is an effective and reliable solution for modern visitor management.