Notes from building voice AI, contact center, and LLM systems in production — written by the engineers doing the work, not marketing.
The TTS voice that sounds best in a vendor demo is not necessarily the one that sounds best over an actual phone line. Here's how to test for the gap before you commit.
Vendor word-error-rate benchmarks are measured on conditions your callers don't have. Here's what actually degrades ASR accuracy in production, and how to test for it before you commit.
A walkthrough of the actual moving parts in a LiveKit voice agent — the pipeline, the room model, and the pitfalls that only show up once real callers hit it.
Both come up in nearly every voice agent build decision. Here's how they actually differ on latency, control, and cost — and when picking the wrong one costs you months.
Vendor pricing pages compare licenses, not total cost. Here's the actual TCO framework — telephony, ASR/TTS per-minute costs, escalation volume, and the maintenance costs nobody puts in the deck.
A practical rollout plan for moving off a legacy IVR to an AI voice agent — what breaks first, how to phase the cutover, and the traps that catch teams by surprise.