AI Order Agent — Ăng ê coffee

Overview

This is an AI ordering and drink-consultant agent that plays virtual front-of-house for Ăng ê coffee — taking orders and recommending drinks so staff stay free during rush hours. I built it solo and rolled it out in phases: starting as a text chatbot, then adding voice, and finally bidirectional real-time voice over HTTP streaming for natural back-and-forth conversation. Reliability and cost were handled with a multi-LLM fallback strategy — Gemini Flash-Lite as the primary model, Groq Llama as fallback — and a self-learning RAG loop (PostgreSQL + pgvector) over the menu that continuously improves recommendations from real interactions. The outcome is a hands-free ordering experience that scales during peaks and gets better over time, reducing staff load without sacrificing the personal touch of a good barista.

Tech stack