Three weeks before the enterprise contract, the voice agent wasn't operator-ready
marcuschen
Command Palette
Search for a command to run...
Archive
20 posts published so far
Your ASR confidence score is a number you can act on. We were throwing it away.
marcuschen
The demo was flawless. The first real call had three people talking.
marcuschen
The call failed on turn nine. My eval gave me one number for the whole call.
marcuschen
Two weeks before launch, every turn was green and the call still died
marcuschen
The guardrail fired at 1.4 seconds. The caller had heard the sentence at 1.1.
marcuschen
Four minutes with the bot, and the human opened with "How can I help you today?"
marcuschen
A month of failed calls, and my eval had the same name for all of them
marcuschen
The voice A/B test that picked the worse agent, and won by 4 points
marcuschen
Our voice agent scored 91 percent. The callers still hung up angry.
marcuschen
The 1.8 seconds after "wait": the week our voice agent refused to stop talking
marcuschen
A caller told our voice agent to ignore its instructions, and it did. The guardrail that fixed it had a 20 millisecond budget.
marcuschen
Ten days before launch, our voice agent kept cutting users off: an end-of-turn detection war story
marcuschen
The caller heard silence for two seconds before the agent spoke
marcuschen
The transcript was perfect and the agent still answered the wrong question
marcuschen
The Friday before we shipped the voice agent, I went looking for 500 callers who did not exist
marcuschen
The transfer was warm on our end. The caller didn't feel it.
marcuschen
The 2am call that dropped before the user finished talking, and the week I spent finding out why my tracer never saw it
marcuschen
End-of-speech detection is the voice-agent setting nobody tunes, and it is why yours feels slow or talks over people
marcuschen
Voice-agent observability is four jobs: 6 tools compared on tracing, eval, guardrails, and closing the loop (2026)
marcuschen