Mmarcuscheninvoicelatency.hashnode.dev·Aug 19 · 7 min readThree weeks before the enterprise contract, the voice agent wasn't operator-readyLook. We had 99.2% uptime in staging. We had eval coverage on 1,400 test turns. We had latency under 280ms first-token. We were not operator-ready. I know this because the enterprise pilot started on 00
Mmarcuscheninvoicelatency.hashnode.dev·Aug 18 · 3 min readYour ASR confidence score is a number you can act on. We were throwing it away.Voice agents treat the transcript as ground truth. The speech recognizer often tells you it is not sure, and we were ignoring it. TL;DR: When I went back through a month of "the agent did the wrong th00
Mmarcuscheninvoicelatency.hashnode.dev·Aug 13 · 8 min readThe demo was flawless. The first real call had three people talking.A scripted voice-agent demo works with one clean speaker. The first production call had crosstalk, and turn-taking fell apart. This is how I fixed it. The demo went perfectly. It always does. One pers00
Mmarcuscheninvoicelatency.hashnode.dev·Aug 13 · 13 min readThe call failed on turn nine. My eval gave me one number for the whole call.The transcript was fourteen turns long and the score was 0.62. That is the entire output. One float, one call, and a rubric that said something like "did the agent resolve the customer's issue." It di00
Mmarcuscheninvoicelatency.hashnode.dev·Aug 10 · 8 min readTwo weeks before launch, every turn was green and the call still diedThe dashboard was a wall of green. Word error rate under 5 percent. Intent classification at 94 percent on our eval set. Response appropriateness, graded by a rubric we trusted, sitting comfortably in00
