AI Voice Agent (Sales Roleplay)
Real-time voice agent that helps sales representatives practice with an AI customer persona.
Approach
It uses a cascaded voice pipeline (ASR → LLM → TTS) being built from scratch to enable natural spoken dialogue. LLM tool calling controls each conversation turn, grounding responses in scenario facts through Qdrant vector retrieval (Vertex AI embeddings). A multimodal LLM-as-judge pipeline scores recorded sales conversations against a versioned rubric.
Evaluation
Both proprietary and open-source speech/language models are being researched and benchmarked to evaluate cost-efficiency and self-hosting feasibility, reducing dependency on third-party APIs.