Make catalogue-aware travel decisions before inventory search.

The expensive reasoning is precomputed into an immutable snapshot, so a request scores a large candidate set against the traveller's constraints in one pass instead of resolving destinations one lookup at a time. Planning returns inside a hard latency budget rather than degrading into a long agent loop.

Ranking is deterministic scoring over versioned data, not generation. A model reads the traveller's sentence; it does not choose the answer. Results change when the snapshot changes and not otherwise, which is what makes them auditable, cacheable, and testable against a fixed benchmark.

Typical conditions by month, seasonality, airport reachability from an origin, points of interest and cost proxies are resolved ahead of the request across tens of thousands of destinations. Widening the candidate set costs compute, not a network round trip or a model call per destination, so lesser-known destinations compete on merit instead of being excluded by a curated shortlist.

Which routes are actually served from an origin, which destinations come in on budget in a given month, which candidates verify successfully and which travellers pursue are carried as signals per destination and month, and are rebuilt into each published snapshot. This is proprietary outcome history that a frontier model does not have unless it is given it.

Budget, dates, duration, origin and party are compiled into a typed contract that records which are binding and which are preferences. Soft constraints are widened only under explicit policy, hard ones are not, and any relaxation is reported back with the result rather than silently absorbed.

Scores decompose into independent signals — intent match, seasonal fit, reachability, experience fit and affordability — so a downstream product can show a reason, a support team can audit a result, and a regression is attributable to a component rather than to a prompt.

Resolve open-ended intent against a managed destination universe while preserving hard traveller constraints.

Reason across flexible windows and trip lengths to prioritise useful combinations instead of forcing one brittle interpretation or shopping every possibility.

Prune, rank and explain the combinations worth live verification, with controlled relaxation when exact options fail.

Plan around caller-known commitments at the supplied destination, with full date coverage by default or any shared local date when requested.

Trip-planning startups, conversational travel assistants, itinerary builders and travel-agent copilots.

OTAs, metasearch, package-holiday companies, flight-search products and publishers adding discovery.

General AI products that need travel discovery capability without rebuilding travel-specific planning infrastructure.