Resort Concierge
2026 — AI Integration / Development / Own Product
AI chat for resort sites. It answers guests in English, Tagalog or Taglish, never invents a price, and hands off to book.
(How it's built)
- 001One script tag adds the bubble to any site. It draws itself inside a shadow root, so the resort's CSS can't break it and it can't break the resort's page.
- 002Each chat goes to a Vercel function. It accepts only known resorts, at most 12 messages of 500 characters ending with the guest's, and browser calls only from listed domains.
- 003The model is gpt-oss-20b on Groq, called through its OpenAI-compatible API. It sees only that resort's facts and answers in the guest's language.
- 004A price guard checks every reply. Any peso amount that isn't in the resort's facts is swapped for the resort's own price answer, so the bot can't invent rates.
- 005With no API key, a quota hit, an outage or a reply slower than 9 seconds, a keyword matcher answers from the same facts. The bubble always replies.
- 006Replies resolve word by word from blur to sharp. Screen readers get the whole reply at once, and reduced motion shows it plain. node:test covers the matcher, the price guard and the API checks.
