How I built trip.expert

Building an AI concierge. Part one: what it does. Developed by Isaac Simantov. What you see is a real recording, not an animation. One tap: check-in and check-out. The answer comes from the hotel's own facts. Or the guest just asks. Is there a swimming pool? The answer comes with photos. One tap, and it fills the screen. Now, someone tries to misuse it. The guardrail stops it, before any AI is called. And when a guest complains, the contact form opens by itself, ready for the hotel. In Hebrew, too. It answers in the guest's language. Fifty of them. Facts, photos, a guardrail, a hand-off to people, and fifty languages. Next: the agentic workflow behind each answer.

Part two: the agentic workflow. One message, nine blocks. Rate limits stop floods before anything runs. Then, cheap checks decide if the turn may run, from session limits to the monthly budget. A pattern filter blocks known attacks before any AI is called. In a single call, the model checks safety, detects the language, and rewrites the question. A router decides the path: small talk goes straight to the answer. This is RAG: vector search finds the matching facts, only from this hotel's knowledge. The model answers under strict rules, and returns text, photos, and whether a human is needed. A second call checks the draft against the facts, and corrects it. Plain code checks the answer: never empty, never too long, never leaking the prompt. Every turn is saved and billed, and unanswered questions go to the hotel's to-do list. Code decides the path. The model decides the words.

Part three: finding the facts. First, the owner teaches the assistant. Then, every guest question uses it. The owner writes plain text about the property, one topic at a time. Long text is split into chunks that overlap, so no fact is cut in half. A language model pulls out every fact, each one a sentence that stands on its own. Each fact becomes a question a guest would really ask, with a short answer. Each pair gets tags, and a vector: a list of numbers that holds its meaning. The new pairs replace the topic's old pairs, all in one step. Now a guest asks: when do we have to leave the room? The question is rewritten to stand on its own, in the guest's language. It becomes a vector too, with the same model. Vector search compares meanings, not words, and only in this hotel's pairs. So a question in Japanese still finds facts written in English. Words like breakfast or airport lift the pairs with matching tags. The model answers only from these facts, and a second call checks the answer. If a fact is missing, the question goes on the owner's to-do list. Teach once. Search by meaning. Answer from facts.