Selected work · 03
Malayalam Phone Reception Agent
A phone receptionist reachable on a real number. It answers as the IoTides receptionist, speaks only Malayalam, answers questions from verified company facts, takes messages, and stores every call's transcript.
Python · LiveKit Agents · LiveKit SIP telephony · Gemini Live API · FastAPI · SQLAlchemy · PostgreSQL · pytest
What it does
A caller dials a real phone number. LiveKit's SIP telephony routes the call into a room, where the agent answers. A single Gemini Live speech-to-speech model does the listening and speaking, so there's no separate speech-to-text → LLM → text-to-speech chain to stitch together.
- CallerDials the provisioned number
- LiveKit SIPDispatch rule creates a call room
- AgentGemini Live, Malayalam only
- FastAPIMessages mid-call, transcript at end
- PostgreSQLCalls, transcript turns, messages
Call path, from the repository's architecture diagram.
Source
The proof of concept was published as an MIT-licensed repository: the agent, telephony provisioning, backend and tests.
This page describes only what's in that repository. My other voice-agent work at IoTides is proprietary and isn't shown here.
What's implemented
- Real telephony: a provisioning script rents a LiveKit phone number and creates the SIP dispatch rule; the agent runs containerised on LiveKit Cloud.
- Agent tools:
get_business_hours,take_messageandend_call. The agent is grounded only in verified company facts. - Barge-in: callers can interrupt; their speech cancels the agent's audio.
- Transcripts on both sides: input and output audio are transcribed and stored with per-turn interruption flags.
- Backend: a FastAPI ingest API with async SQLAlchemy models for call sessions, transcript items and caller messages.
Tuned for phone audio
The turn-detection settings were chosen from real call logs:
| Setting | Value | Why |
|---|---|---|
| End-of-turn silence | 400 ms | Reply quickly without cutting callers off |
| Speech sensitivity | High (start and end) | Catch interruptions on narrow phone audio |
| Prefix padding | 300 ms | Keep the start of each utterance |
| Model thinking budget | 0 | Removes multi-second variation in reply time |
Testing
The repository README lists 23 pytest tests in three layers. I haven't re-run them for this page; they're the repository's own claims at publication.
- Unit tests: session-report to transcript-payload conversion, with no services needed.
- Integration tests: the FastAPI API and an end-to-end transcript flow against PostgreSQL.
- Behaviour tests: the real agent run through the LiveKit Agents test framework, judged by an LLM.
Known limitations
- It's a proof of concept, not a production deployment.
- No transfer to a human. The agent can take a message or end the call, but it doesn't hand calls over.
- Cloud transcript storage isn't wired up. Until the backend is hosted somewhere reachable, the cloud agent saves transcripts as JSON files inside its container.
- One language, one persona. It speaks only Malayalam, by design.