How are you reliably handling spoken dates, times, and addresses in your LLM + voice AI flows? Especially when integrating APIs — what's working for you?