DINA
Voice agents for inbound and outbound calls, with a pipeline any agent can plug into as the brain.
- Where
- At Decklar
- Status
- In production
- My part
- Worked on the realtime voice pipeline, the voice experience and latency, and making the agent behind it pluggable.
- Built with
- LiveKit, Deepgram, OpenAI, Twilio
DINA is Decklar’s voice agent. It handles inbound and outbound calls, collects information from the caller, and can start agents and workflows while the call is still going.
The pipeline uses Deepgram for speech to text, OpenAI for responses, and LiveKit and Twilio for telephony. It started as a proof of concept, was used for customer demos and onboarding, and is now in production.
Latency
Speech recognition, the model, any tool the agent calls, and speech synthesis all add to the pause before the caller hears an answer. A lot of my work is managing that chain so the call feels like a conversation.
One pipeline, any agent
I made the brain pluggable. The pipeline owns the call itself: audio, speech to text, turn-taking, and speech back out. The agent that decides what to say plugs into it.
So different agents can run a phone call without each one rebuilding the voice layer, and improving the pipeline improves all of them.