2025 · Airbase

Real-time Voice AI Interview Agent

An AI agent that conducts live job interviews over the web and phone, in real time.

TypeScriptWebSocketsTwilio Media StreamsSTT/TTS/VADBun

The role

Lead Engineer — architecture & implementation

Highlights

  • Built a sub-second, barge-in-capable voice pipeline bridging browser and phone callers into one agent.
  • Designed a resilient WebSocket transport with client-side recording and Service-Worker upload reliability.
  • Replaced a costly managed voice stack with a custom bridge, cutting per-minute cost dramatically.

The problem

An AI interviewer only works if it feels live — it has to listen, think, and respond fast enough that the conversation never feels robotic, on both a laptop mic and a phone line.

What I built

The real-time voice system inside Airbase. Audio streams in over WebSockets (browser) and Twilio Media Streams (phone), runs through a speech-to-text → LLM → text-to-speech pipeline, and streams the reply back — with barge-in so the candidate can interrupt naturally.

Every hop is measured to keep responses under a second; client-side recording plus a Service Worker means no lost audio on flaky networks; and swapping a managed pipeline for a custom bridge cut the per-minute bill by an order of magnitude.

There's a longer write-up in the blog.

Want something like this built?

~/work/voice-interview-agent$