Work logs, teardowns, and negative results from the CallingScout voice platform team. Most of this is what we wish someone else had written before we started.
Most platforms bolt the legal disclosure onto the system prompt and hope the model says it. We inject it at the pipeline factory so the first utterance exists before any LLM token is generated. The hard part isn't the injection. It's the warmth.
A natural voice turn has about 600ms of silence before it starts feeling like a bad connection. That's the budget. Here's how we spend it across the pipeline, and where we cheat with speculative warm-starts.
Every backchannel (mm-hmm, uh-huh, yeah) should encourage the agent, not cut it off. Default voice-activity detectors treat them as interruptions. We wrote a backchannel classifier that lives upstream of the aggregator and runs in 4ms.
We burned a week chasing why mobile-hotspot test calls would connect, ICE-negotiate, then die after 12 seconds. This worklog walks through the ICE restart rigging, the 20s/40s heartbeat ping-pong, and why the STUN-only fallback was silently lying to us.
Do-not-call registry lookup, TCPA calling-hour timezone math, and two-party consent disclosure all live inside one function. If a dispatcher bypasses it, the outbound dial literally doesn't happen. $43,792 per violation is a useful forcing function.
We ripped out our scripted-flow library two weeks after adopting it. Language-driven slot filling plus a single typed tool call per slot gives the same determinism, ten times the resilience, and zero diagram-maintenance tax.