I played around with livekit last week and put a public phone number on my agent and made it show its own call flight data. This was just to play as I’m looking for solutions for live voice customer assistance that I’m building.
Every call’s turns are written live into EOU detection delay, LLM TTFT and TTS TTFB against a 1s response budget, with interruptions, running cost, etc. I also made a “chaos mode” that injects latency so you can watch the instruments catch it.
Demo: cockpit.agentmaker.io (browser call, a SIP number, or a recorded sample flight if you don’t want to grant mic). Fleet-level view (p50/p95, over-budget rate, degraded turns, spend) is on a public Grafana dashboard: Grafana
Repo on footer.
The whole cockpit is built on the stock metrics events from Agents + Inference. That layer is actually really excellent and he per-speech-id events were rich enough that I never had to instrument anything myself.
Curious what others are seeing in the wild: my last 14 days on Inference are p50 1.31s / p95 3.56s against a 1s budget, with 64.6% of turns over budget. What p50s are you holding on public traffic, and with which stack?