Research and MVP Development of an Inbound and Outbound Voice Agent for Dewly AI, Mapped from the Existing Agentic Chat System

Dewly AI is a LangGraph-based multi-agent hotel-booking assistant, originally built for text chat: a router dispatches each message through guardrails and intent classification to one of four domain agents, before a synthesizer composes the final reply. This capstone project extends Dewly with a second channel, voice, wrapping the same agent graph with Deepgram STT, Cartesia TTS, and LiveKit's real-time audio pipeline to support live, phone-style conversations rather than just text exchange. Voice introduces failure modes chat never faced: silent dead-air, mishandled interruptions and backchannels, missing filler phrases during slow lookups, and imperfect turn segmentation. To catch these, the project also built EVA-for-Dewly, an evaluation framework that drives real LiveKit conversations against Dewly's voice channel - either via a scripted LLM persona (bot-to-bot) or a human caller (human-to-bot) - producing transcripts, audio, and turn timestamps. Dewly is treated as a black-box system under test, scored by a mix of rule-based metrics and LLM-judged/LALM-judged metrics, aggregated per run. The project is conducted under academic and industry supervision, with Dewly AI's CTO as industry supervisor.