Verve Athletics Voice Support Agent

Verve Athletics Voice Support Agent is a full voice conversation loop built for the AssemblyAI Voice Agent Hackathon. A customer speaks a question, it's transcribed in real time by AssemblyAI, answered by a Gemini-powered RAG pipeline grounded in real store policy documents, and spoken back out loud through ElevenLabs. No typing, no forms, just a natural conversation. The system handles real-world complexity beyond simple Q&A: multi-turn conversations for order verification (collecting an order number, then a total, to confirm identity), per-session rate limiting to prevent runaway API usage, and a concurrency guard so a second question mid-answer doesn't clobber the first. The standout feature is conflict detection: if two retrieved policy chunks genuinely contradict each other on the same point, the agent doesn't blend them into a confident-sounding guess. It flags the discrepancy and escalates to a human instead. During development, this caught a real, previously unnoticed contradiction between two policy documents, not a scripted demo case, but the system finding an actual error in its own knowledge base. Built on AssemblyAI for real-time streaming speech-to-text, Google Gemini for RAG-based answer generation, ElevenLabs for natural text-to-speech, and FAISS for policy retrieval. Deployed as a genuine production system, Flask-SocketIO backend on gunicorn with threaded workers, behind nginx, managed by systemd, secured with Let's Encrypt SSL on AWS EC2. Try it live at verveassist.live.