Tartendu Kumar AI Systems Engineer
00Index 01Work 02About 03Log 04Press 05Contact 06Résumé

Index / Work / AI Voice Agent with Memory

Voice AI Hospitality · Automotive · Healthcare Web

AI Voice Agent with Memory

Sole architect and developer — architecture through deployment

A production voice agent that answers and places real phone calls in 90+ languages, remembers who it is talking to, and acts on what it hears. Realtime audio transport, session management, retrieval and tool execution — built and deployed solo, and running in production for multiple operators.

FastAPIasyncioWebSocketTwilioFirestoreRAG
AI Voice Agent with Memory
01 Specification
Client
Zencia.ai
Industry
Hospitality · Automotive · Healthcare
Role
Sole architect and developer — architecture through deployment
Platform
Web
Stack
FastAPI · asyncio · WebSocket · Twilio · Firestore · RAG
Dated
2026-04-08
02 Key features
  • Realtime speech-to-speech with barge-in — callers can interrupt mid-sentence
  • Persistent per-caller memory injected into context at call setup
  • Async session registry: 50 concurrent calls, idle reaper, reconnect with exponential backoff
  • Bidirectional PCM audio streaming, 16 kHz in / 24 kHz out
  • Retrieval over per-agent knowledge bases for grounded, non-hallucinated answers
  • 14 agent tools — calendar booking, CRM writes, warm transfer, messaging
  • Post-call transcription, deduplication, summarisation and automated follow-up
  • Three telephony providers behind one interface
  • Cloud or fully on-premise deployment, with a zero-cloud-dependency option
04 Problem → solution → outcome

① Challenge

Voice agents fall apart in production for three reasons: they forget the caller the moment the line drops, they talk over people, and once the call ends they cannot actually do anything. A demo hides all three. A real deployment exposes them on day one.

② Solution

A realtime speech-to-speech pipeline rather than the usual transcribe–think–speak cascade, so the agent hears and answers natively with barge-in handling — a caller can cut in mid-sentence and it stops. Persistent per-caller memory is injected into context at call setup, so a returning caller is recognised rather than re-interrogated. An async session registry holds 50 concurrent calls with an idle reaper and reconnect-with-backoff. Mid-call the agent queries a per-agent vector store for grounded answers and invokes tools to book, write to CRM, and hand off to a human.

③ Result

Live in production across multiple operators, handling inbound and outbound calls in 90+ languages. After each call the system transcribes, deduplicates, summarises and fires follow-ups over SMS and WhatsApp without a human touching it. The same stack deploys to cloud or fully on-premise, where no audio ever leaves the device.

Need something
like this built?