Engineering

How voice AI systems detect whether a human, voicemail, or IVR answered the phone, and how to handle each scenario appropriately.

How to preprocess text for speech synthesis so phone numbers, postcodes, dates, and business data are pronounced correctly.

How voice AI can autonomously navigate phone menus using DTMF tones while preserving conversation state across connection interruptions.

How to preserve conversation context when WebSocket connections drop and reconnect in real-time voice systems.

How pre-generated audio acknowledgements can make your voice AI feel instant, even when processing takes time.

A deep dive into debugging real-time transcription latency issues caused by misunderstood event routing in speech-to-text APIs.

What we learned building SwiftCase's workflow engine over 15 years - from modular architecture to event-driven processing. The decisions that let us scale to millions of cases.

How Switchboard combines custom-built AI models with external providers - OpenAI, Anthropic, Deepgram, ElevenLabs - to optimise for each task while avoiding vendor lock-in.

Building voice AI agents that feel natural is harder than it looks. Here are the technical challenges we solved building Switchboard - from interruption handling to turn-taking.

SwiftCase has been in production for 15 years. Here's what we learned about building software that lasts - from technology choices to team practices.