Realtime speech, both directions
Streaming speech-to-text and text-to-speech wired so the agent listens and talks at conversational latency — the delay stays short enough that the back-and-forth feels like a call, not a walkie-talkie.
Realtime voice agent· Live
AI call assistant
Not a phone menu. A voice agent that picks up a real call, follows the conversation wherever it goes, and actually does something about it.
Realtime AIVoice / STT / TTSLLM agentsTelephonyNodeTypeScript
The challenge
What we were up against.
Phone lines are where a lot of businesses lose people — calls go unanswered, or callers get trapped in a press-one-for-this tree that never fits what they actually want. The bar for a real alternative is high: an agent has to understand natural speech in real time, respond fast enough that the pause doesn't feel wrong, and take a real action at the end instead of just talking.
What we built
The parts that mattered.
Understanding, not keyword matching
An LLM interprets what the caller actually means across a winding conversation, handling interruptions and changes of direction instead of forcing them down a fixed script.
Agent that takes action
The assistant is wired to tools and backend actions, so a call ends in something done — a booking, a lookup, a record updated — rather than a promise to pass the message along.
Telephony integration
Connected to the real phone network so it answers actual inbound calls, not a demo in a browser tab.
The outcome
Where it landed.
It's live and handling real phone conversations end to end — understanding the caller, replying naturally and completing the task on the line. It's the difference between a business missing a call and a business quietly answering it.
Services behind this