An AI voice assistant for pilot customer calls

Context

A technical prototype exploring whether automated Czech-language calls could reconnect with prospective customers and schedule a technician meeting through a natural conversation.

My contribution

• 01Implement scheduled call initiation and connect live Twilio calls to the AI assistant.
• 02Pass transcriptions and customer context to the LLM, and co-author tools for proposing dates, saving appointments, and ending calls.
• 03Persist conversation state and meeting results while supporting multi-turn Czech dialogue.

Decisions that shaped the work

• 01Use a dedicated speech-recognition service that handled Czech more reliably than the available model-native recognition at the time.
• 02Expose a small set of scheduling actions as tools so conversations could produce structured outcomes rather than dialogue alone.
• 03Handle interruptions by stopping assistant playback and resuming from the preserved conversation state—the hardest part of making calls feel responsive.