How I built an AI email agent that processes 15,000 hotel guest emails per day. full architecture breakdown

Just shipped this project and wanted to share the full technical breakdown because hotel/hospitality AI doesn't get much attention compared to the usual chatbot and SaaS use cases. The client manages 500 hotel properties. Their support team was manually handling around 15,000 guest emails per day. Same questions over and over across hundreds of hotels but each one still needed a human to read it, understand it, find the answer, and reply. Here's how the system works end to end: Layer 1: Email ingestion and question extraction This was the hardest part. Guest emails are messy. A typical one looks like: "Hi there, we're coming for our anniversary on the 20th and I was wondering if you have any room upgrades available. Also is the spa open to guests or do we need to book separately? We're driving so need to know about parking too. Last time we stayed the wifi was a bit slow in our room, has that been fixed? Thanks!" That's four separate questions plus a complaint wrapped in one email. If you just embed the whole thing and search the FAQ database you get a blended result that partially answers one or two questions and misses the rest. So I built an extraction layer that reads the full email and breaks it into individual questions. It handles directly stated questions ("is the spa open?"), implied questions ("we're driving" implies they need parking info), complaints that need acknowledgment but aren't FAQ-searchable ("wifi was slow"), and informational context that shouldn't be treated as a question at all ("coming on the 20th"). Getting this extraction reliable was probably 40% of the total development time. Layer 2: FAQ knowledge base with vector search All hotel FAQs get embedded and stored in a vector database. Different properties have different amenities, policies, and details so the search is scoped per hotel. When a guest emails the Berlin property asking about breakfast, it searches the Berlin FAQ, not the Munich one. Each extracted question from Layer 1 gets s