Room mapping for a hotel marketplace, live and watched for drift.

Room mapping decides which supplier rates belong to the same physical room. A hotel sold through several suppliers arrives as many names for a few rooms, with different wording, bed descriptions and views. If the names are not grouped, a marketplace shows the same room several times; if they are grouped too eagerly, it promises a room the guest will not get. The monitor tracks exactly those two failures, as under-promise and over-promise, alongside the purity of each group.

A tenant sends one hotel's rates to a single endpoint, up to 600 per request, in standard or consolidated mode. The model groups on the room name; a strict merge policy also reads beds, view and description before joining two rates. Each group comes back with the suppliers' images, descriptions, amenities and occupancy merged and attributed, plus a policy version, so any result can be traced to the rules that produced it.

Partners report a wrong merge through a feedback endpoint. The report becomes a cannot-link applied on every later call for that key and hotel, so the same mistake does not come back. Uncertain results go to a low-confidence review queue, which exports model-proposed and human-approved labels separately. In production the model maps names with a deterministic canonical-key vocabulary and an embeddings fallback; the language model was taken out of the hot path.

A separate monitor watches the marketplace's production model. A health check reads the last six hours of inference logs from AWS CloudWatch, compares them with a baseline taken on deploy day, raises drift alerts and lists room-name patterns the model has not seen before. A dashboard shows drift, group purity, over-promise and under-promise, and sets Plexe's groupings beside those of the incumbent mapping vendor.

Fixes follow a written runbook: classify the gap, measure its blast radius across a corpus of about 32,000 production room names, bump the model version, deploy, verify, and update the watch. Results are cached in S3 for seven days, keyed by the hotel and its set of room names, so a supplier that changes its names gets a fresh mapping rather than a stale one.

The API and portal are live in production at sandbox.plexe.ai, with per-tenant quotas, an attribute guard, version pins for rollback and an uptime probe every ten minutes. Each release is judged on live traffic by counting the wrong merges it removes and the false splits it introduces, and the release notes publish both numbers.

Batch mapping runs as a bounded pilot: up to 1,000 hotels or 32 MiB per job, results kept for seven days, and a resumable client for larger catalogues. Tenants submit whole hotels, never one hotel split by supplier, and each accepted hotel counts once against the daily quota. The pilot claims no accuracy figure yet.