Sangati is developing a local, privacy-conscious observation layer for restaurant floors. Its Mumbai pilot is currently validating person and table occupancy, observed staff table checks, and staff-recorded Ready status. There are no published, independently verified performance or revenue outcomes.
1. Occupancy is an observation, not a service outcome
A restaurant's POS knows which bills are open. Its floor team knows what is happening in the room. Camera-based occupancy aims to add a third kind of information: whether people are observed in the mapped seating area at a particular time.
A credible system needs to distinguish the table from the chairs and nearby walkways. Merely detecting a person in a rectangle around a table can count passers-by, staff and neighbouring diners incorrectly. Camera position, overlap, occlusion and seating configuration all matter.
Even correctly observing an occupied table says nothing about the quality of service. That distinction matters operationally: “occupied for 12 minutes” and “waiting 12 minutes for an order” are different statements.
2. What Sangati's pilot is testing
| Signal | What it means | What it does not establish |
|---|---|---|
| Person detection | People are observed within an assessed camera view. | Identity, emotion, intent or a guest profile. |
| Table occupancy | A mapped seating area is estimated occupied or unoccupied. | Whether guests have ordered, paid or are satisfied. |
| Staff table check | An observed visit is recorded, including time since the last visit. | What was said, delivered or resolved. |
| Staff-recorded Ready | A team member marks a table Ready in the workflow. | Camera proof of cleaning or reset completion. |
Validation status: These functions are being field-checked in Mumbai. Greetings, dish delivery, bill intent, kitchen bottlenecks and measurable service uplift are not validated capabilities of the current pilot.
3. Where false alerts originate
Restaurant dining rooms are difficult computer-vision environments because people move unpredictably and tables are partially concealed. A promising demonstration can fail in a real lunch rush. A serious evaluation should inspect:
- Wrong table mapping: the detection zone covers the tabletop but not the actual chairs or seating positions.
- Occlusion: another diner, partition, waiter or decorative object hides the seated person.
- Walk-throughs: staff or guests pass near a table and trigger a mistaken occupancy or check event.
- Lighting changes: shadows, reflections, evening service and camera exposure affect detections.
- Operational ambiguity: a staff member standing beside a table may not be serving that table.
- Infrastructure: camera streams, recorder compatibility and local device capacity can affect continuity.
The appropriate response is not to claim perfect detection. It is to define conservative thresholds, collect representative observations and quantify both missed events and unnecessary alerts before relying on notifications.
4. What a credible restaurant pilot should publish
Before a venue treats occupancy or staff-check alerts as an operational control, the following measures should be evaluated against manually verified floor observations:
- Camera and table coverage: which seating zones are observable, and which cannot be assessed reliably.
- Occupancy precision and recall: how often “occupied” detections are right, and how often genuinely occupied tables are missed.
- False alerts per shift: not just technical accuracy, but nuisance incidents that interrupt service.
- Check-event reliability: whether staff visits were genuinely observed, and how long the system took to reflect them.
- Busy-period performance: verify performance during rushes, not only quiet periods.
- Workflow separation: distinguish camera observations from employee-entered statuses.
Sangati has not yet published a verified scorecard for these measures. Any future case study should state the sample size, definition of a correct result and limitations.
5. Privacy and practical deployment
Sangati's stated design uses person-only detection without facial recognition, identity inference, guest profiling or audio analysis. Camera inference takes place on assessed venue hardware. Authorized remote dashboard access is a separate layer and may make operational records accessible outside the venue.
Existing cameras are not automatically compatible. The venue must assess RTSP or recorder access, camera angle, mapped seating coverage, processor capacity and competing workloads. Local inference does not remove the need for agreed retention and access controls.
Read Sangati's privacy posture and the product workflow for current details.
Frequently asked questions
Can Sangati work with existing restaurant CCTV?
Potentially. Compatibility depends on camera access, recorder support, view quality, table coverage and a suitable local processing device. Suitability is assessed venue by venue.
Does the camera know that a customer wants the bill?
No. The current field pilot does not validate bill-intent detection. Occupancy and staff visits should not be presented as proof of guest intent.
Does Sangati recognise guests?
No. The stated pilot design does not use facial recognition, identity inference or guest profiling.
Does occupancy AI replace a restaurant manager?
No. It is intended to give staff more timely floor observations. Managers still own context, prioritisation and service decisions.
Is Sangati proven to increase restaurant revenue?
Not at present. The field pilot has not established revenue effects, guest satisfaction improvements or measured service uplifts.
Evaluate your floor, not a marketing promise.
Interested in testing whether your existing camera coverage is suitable? Sangati is accepting early-access enquiries and will assess the venue before suggesting deployment.
Request a venue assessment →Published by Sangati · Mumbai, India · Updated 18 October 2026. This page explains the current pilot scope and evaluation criteria, not independently audited results.