AI voice assistant. A real-time voice assistant, from demo to production in two months.
talks
The Fifth Elephant
Lessons from building real-time voice AI applications
We built an AI voice assistant for a leading consumer technology platform with 1M+ app downloads.
We shipped the AI voice assistant from a one-day demo to handling live calls with 1,000 internal users in two months.
We owned the real-time voice stack end to end: bidirectional audio streaming, voice activity detection, barge-in, end-of-speech detection, multilingual transcription, intent classification, and live and post-call summarization.
We handled hard conversation problems in production, including interruptions, human-versus-robot caller detection, and mid-call goal changes driven by user input.
We also set up the evaluation framework and observability needed to measure call quality, debug failures, and keep improving the assistant after launch.