Designed around
the working kitchen.

Start with a site assessment.

We check camera compatibility, network access, lighting and the view of the agreed work areas. Existing cameras may be suitable; compatibility is not universal. Camera movement requires the zones and measurements to be checked again.

The pilot starts with a small number of cameras and clearly defined standing areas. The restaurant supplies its service hours and timezone, nominates a manager and arranges appropriate permissions and staff communication.

Privacy and data flow

Privacy processing begins on site.

1. The edge box receives the camera feed. Face masking and configured redaction areas are applied locally. The intended pipeline sends processed images onward for analysis.
2. Cloud processing produces measurements. Person detection and work-area measurements currently run in the cloud. This is not an entirely on-device analytics product.
3. The manager receives observations. Reports describe agreed work areas. They do not identify staff or provide individual performance scores.

Face masking can fail. It is not a guarantee of anonymity, and removing faces alone does not remove every identifying detail. Privacy validation, image-path checks and an agreed operating configuration are prerequisites for customer activation.

Pilot data access, retention, permitted uses and deletion are agreed in writing. Training-data use requires a separate explicit arrangement; agreeing to a pilot is not blanket permission to train on kitchen footage. No audio is collected in the pilot pipeline.

What we measure

Observations with context.

Presence in agreed work areas

Time-weighted estimates of how many people were observed in an area. Gaps and ambiguous assignments must not be treated as an empty kitchen.

Time spent in an area

Observed visit durations, subject to camera visibility and tracking limitations. Time at a counter is not a measurement of productive work or a specific task.

Coverage

How much of the configured observation period was actually measured. Without configured service hours, an observed-window statistic is not a full-service coverage claim.

Comparison with prior observations

When enough eligible history exists, we compare matching time buckets for the same configured area. A difference is a prompt to investigate, not a diagnosis.

What the 21-day baseline means.

The current baseline looks back over a rolling 21-day window using 15-minute time buckets, with weekdays and weekends grouped separately. It needs at least five eligible daily observations for a matching bucket. It is not an exact Friday-to-Friday comparison.

Twenty-one days is not a model-training period or a guaranteed waiting time. Missing coverage and changed camera or zone configurations can delay comparisons. Reports can show validated observations earlier, clearly marked when comparative history is insufficient.

What the pilot does not promise.

No facial identification, employee ranking, handwashing or PPE certification, guaranteed savings, or conclusions about why service was slow. Without order-system data, we cannot infer ticket times or sales from presence alone.

Commissioning includes privacy and timed accuracy tests. Availability, failure reporting and recovery are checked before activation. Our current programme is pilot-stage; the scope and acceptance criteria are agreed before installation.