Skip to content

A voice agent that actually listens.

Maya is Drevan's voice agent for sales, support and appointment requests. One agent. Many roles.

Talk to Maya now.

Sign in with Google and talk to Maya.

If the demo server is offline, the live page falls back to text mode.

Book a live demo
50 s launch film. It opens on the problem Maya is built for.

Video not loading? Try Maya live (opens in new tab) or book a live demo.

What Maya does.

Sales, support, appointments.

One agent covers all three roles.

  • Interruptions and changes

    Handles interruptions and mid-call changes of mind.

  • Live transcript

    Every word shows as the call happens.

  • Post-call brief

    What the caller wanted, what was captured, what was missed.

  • English, Hindi and Arabic

    Three languages, in regional styles.

What Maya is not

Maya says she is an AI on every call. Her five demo businesses are fictional.

Seven stops, every one of them on the clock.

Sarvam Saaras v3 speech recognition, then a Groq-hosted open-weight 120B model, then Sarvam Bulbul v3 speech, over LiveKit WebRTC, with per-call cost tracing.

Signal chain: Caller (Phone), VAD (Silero), STT (Sarvam), Substrate (Drevan), LLM (Groq), TTS (Bulbul), Transport (LiveKit). Substrate is the only Drevan node.

  1. Silero VAD

    Detects speech boundaries locally, so the network is never the thing deciding whether someone is talking.

  2. Sarvam STT

    Streaming transcription tuned for Indian languages and code-switching between Hindi and English mid-sentence.

  3. Groq LLM

    Inference chosen for latency, not benchmark scores. A correct answer that arrives late is a failed turn.

  4. Sarvam Bulbul TTS

    Streaming synthesis. Audio starts before the sentence is finished generating.

  5. LiveKit WebRTC

    Real transport with real jitter buffers, because a phone call is not a websocket.

What is actually running.

  • Silero VAD
  • Sarvam STT/TTS
  • Groq
  • LiveKit WebRTC
  • LangGraph
  • vertical config packs

What real calls broke that testing never would.

Barge-in that cancelled itself

Interrupt handling wedged the pipeline when the agent was cancelled mid-utterance. Found by listening to real recorded calls, not by testing. Fixed: Maya now stops when the caller speaks and answers what they said.

False-interruption resume

A cough or background noise read as an interruption and killed the sentence. The agent now resumes rather than abandoning the turn.

Ringback transcribed as speech

The dial tone itself entered the transcriber and became input before the human ever spoke.

Turn latency, decomposed

Per-turn latency is broken down by stage in the call report, so a slow turn names the stage that caused it instead of being "the agent felt slow".