An AI concierge that learns each member's taste.
Two Gemini agents that search and hold real hotel rooms, a feed that learns from every swipe, and a taste memory the member confirms and controls.
Build the intelligence behind a travel app: an assistant that can actually check and hold rooms, and a feed that gets better for each member without ever booking, charging or remembering anything on its own.
A Python backend with a Gemini concierge and reservations agent, an offline Gemini tagger and embeddings for every hotel, a per-member learning feed with exploration and taste neighbours, a confirmed-only taste memory, and a booking flow where a person always signs off.
Each feature below is mapped to the part of the system that runs it: Gemini, a machine-learning model, fixed rules or a person.
Gemini 2.5 Flash on Google ADK: a concierge agent hands bookings to a reservations sub-agent with tools for live rates, availability, dates, currency and the member's profile. It replies in the language the member writes in.
Gemini reads each hotel's catalogue text and scores it on 19 taste dimensions and 9 practical facts, with one line of evidence per score. No signal in the text means no score, not a guess.
Gemini embeddings place every hotel in a taste space, plus a style view with the city's average removed, so the feed matches feel, not just location.
Each member has their own Bayesian model, updated from dwell time, photo swipes, saves, bookings and a reason for 'not for me'. Exploration tries new options, preferences fade over 120 days and a shift in taste is detected.
Collaborative filtering finds members with similar taste, a diversity re-rank stops the feed repeating itself, and each card shows the room that best fits the member.
A position-debiased ranker trained on feed logs runs in an A/B arm against hand-set weights, with off-policy evaluation. Hand-set weights serve by default, and tests so far use simulated members.
The model may quote a price only after a tool returned it. Its replies are checked for personal data and prompt leaks, and dealbreakers such as step-free access filter results and never fade.
The agent cannot confirm, charge or cancel: a person verifies the hold and the member confirms with their own button. A taste fact is saved only after the member says yes, and one call erases what the feed learned.
Card details go only into the hotel partner's hosted card form via a signed link, phones are linked with Twilio Verify SMS, and partner webhooks keep trips in sync.
Will connect the app screens (assessment, archetype and Ask Loma) to this engine. Today they are a designed front end, not yet wired to it.
Gemini agents that search, compare and hold rooms in the member's language.
Hotels ranked per member from what they open, save, skip and book.
Confirmed preferences kept per trip type, with dealbreakers that never fade.
The room on each card is picked for the member, not just the hotel.
Partner-hosted card form, human sign-off and the member's own confirm button.
Bookings kept in sync with the hotel partner by webhook.
Know what it will cost and what it will return before you build it.