Kiishi.AI Front Desk is a real-time voice receptionist that answers calls, manages appointments, remembers returning callers, and automates follow-ups through Google Calendar and other business tools.
Small businesses can lose potential customers when calls go unanswered, especially outside working hours. Hiring someone to answer every call can also be expensive. I wanted to build a voice receptionist that could handle common requests without requiring someone to stay on the phone. The engineering challenge was making the conversation feel natural while connecting it to actual business systems. The assistant needed to recognize callers, understand appointment requests, check calendar availability, and complete bookings without getting dates wrong. Response time was another concern. A voice assistant that takes too long to answer creates uncomfortable pauses, so reducing latency became an important part of the implementation.
I developed Kiishi.AI Front Desk, a real-time voice receptionist built with Python, FastAPI, and Pipecat. The system uses Telnyx for phone calls, Speechmatics for speech recognition, Anthropic Claude Haiku for conversation handling, and ElevenLabs for voice generation. It connects to PostgreSQL to retrieve caller information and Google Calendar to check availability and manage appointments. I implemented ten tools for actions such as booking, rescheduling, retrieving context, and ending calls. I added a custom pronunciation dictionary with 1,000 entries to help Speechmatics recognize African names. A separate date-resolution tool handles calendar calculations. To reduce response time, the application prepares caller data, calendar connections, and voice services while the phone is ringing. After each call, GPT-5-nano extracts a structured summary. Webhooks send the results to n8n for follow-up automation. The system records itemized call costs across five services in PostgreSQL. I deployed it on AWS EC2 using Docker, GitHub Actions, AWS Systems Manager, and OIDC. CloudWatch provides error monitoring.
Kiishi.AI Front Desk demonstrates how incoming calls and appointment-related tasks can be handled through an automated voice workflow. The receptionist can answer calls, retrieve information about returning callers, check calendar availability, book or reschedule appointments, and trigger follow-up actions. I reduced the delay before the assistant's first response by preparing services during call setup. The system can deliver its first response in under three seconds. I also added a pronunciation dictionary to help with African names and a deterministic date resolver to reduce calendar calculation errors. Every call has an itemized estimated cost record covering the services used. CloudWatch logs and deployment automation make the application easier to maintain and investigate when problems occur. The application has been deployed on AWS and supports live phone interactions. I have not yet established verified customer savings, appointment conversion rates, or reductions in staffing costs.
The deployed voice receptionist achieves a first-response time of under three seconds through service preparation during call setup. The implementation includes ten callable tools for handling appointments and caller interactions, plus a 1,000-entry pronunciation dictionary for African names. An itemized cost-tracking system records estimated call costs across five services: telephony, speech recognition, speech generation, conversation processing, and post-call summarization. These are implementation and operational metrics. Customer-level time savings, booking success rates, and financial savings have not yet been independently measured.