Openline

Guests message a hotel’s WhatsApp number. The system retrieves that hotel’s own knowledge, drafts a reply with a locally hosted language model, and either sends it or holds it for a human to approve from an iOS app or a web console. The models, the database and the WhatsApp session all run on one machine; only the API is reachable from the internet.

The Problem

A hotel front desk answers the same forty questions every day — check-in times, parking, pets, late checkout. A language model can answer those. What it must never do is invent a price, promise a refund or improvise a reply to a complaint.

The Solution

The product is built around one rule: the AI drafts, a human decides. Retrieval scoped to each hotel, a policy engine, a review queue and a full audit trail exist to make that decision fast, informed and reversible.

What I Built

• →Answers from the hotel’s own knowledge — mined from past conversations only where an admin selects them
• →Time-aware retrieval: an old price is flagged and ranked below current facts
• →Refunds, discounts, pricing, complaints, legal and payment topics always require a human
• →Each draft records model, intent, confidence, policy reasons and every retrieved source
• →A new guest message supersedes a waiting draft instead of sending a stale reply
• →Approvals, edits and declines become training signal for the next answer

Highlights

• →Local LLM — nothing leaves the machine
• →Policy floor: money and complaints always go to a human
• →Every draft explains its sources

Outcome

• →Runs end to end on one Mac Studio: language model, database, queue and WhatsApp session
• →Only the API is exposed, through a Cloudflare Tunnel — Postgres, Redis and the model are never public
• →Operators approve from an iOS app or a web console with push and live updates