Smart Table Service Intelligence

The simulated camera samples a top-down frame for a table on a gentle cadence rather than streaming. If the frame is effectively identical to the last one, the change-gate stops right there - the last perception is reused and nothing is spent. This is the first and most important decision in the journey: most frames don't change, so most frames cost nothing.

When a frame has changed, the one metered vision stage reads it: the tableware present, each item's fill state, and a confidence. Tracking ties those detections to persistent item identities across frames - so it knows it's the same glass, not a new one - and aggregation rolls the items up into the table's condition and dining phase: seated, dining, finished or idle.

A salient situation - drinks empty, plates finished, an anomaly - must repeat across N consecutive sampled frames (default three) before it counts. This confirmation window is the false-alarm guard: a reflection, a passing hand or a single misread frame never survives it. Only a confirmed situation reaches the rule layer, which maps it to a specific action and priority via the two-layer baseline-plus-overlay rules.

Governance suppresses duplicates of an already-open alert within a cooldown window, then the alert is dispatched to the assigned server with its snapshot. From there the lifecycle runs: acknowledge, resolve, escalate on timeout, confirm the resolution from later frames. And much of the journey is the system staying quiet - a steady table, an unconfirmed flicker, a degraded frame routed to manual check. Those non-events are the product working as intended.