Sentinel — bar-queue & ambient retail watch
Edge cameras + small object-detection model giving bar staff a live read on queue length, refusal risk and crowd density.
Pavilion's bar team had no quantitative signal of queue pressure on busy nights. Decisions on pulling extra staff or opening a second till were made by whoever happened to look up. By the time someone noticed a five-deep queue, walk-aways had already happened.
Deployed cheap PoE cameras above each bar pointed straight down. A small YOLO-class detector runs at the edge counting heads inside fixed polygons. Inference happens on a Mac mini, not in the cloud — no images leave the building. Aggregated counts get pushed every 10s to a dashboard the head bartender keeps on a tablet, with a Dispatch alert when the threshold trips.
Object detection + occupancy estimation. Nothing fancier — no tracking, no faces, no behavioural inference.
Defined the trigger thresholds, designed the operational response, trained staff on what the alerts mean, and stayed the decision-maker on staffing.
Bar refusals dropped sharply on the busiest 6 nights of pilot. Just as importantly, the head bartender stopped having to physically walk the floor to assess pressure.
First model was trained on stock pedestrian data and missed people sitting on stools as 'in queue'. Re-labelling 400 frames from the venue itself fixed it in a day. Lesson: site-specific data trumps bigger general models.
Keep the AI doing one narrow thing it's verifiably good at, and route the decision to a human who already had authority. The dashboard isn't an autopilot — it's a better window.
- 01Edge inference removes the GDPR/privacy conversation almost entirely — no images leave the premises.
- 02Don't try to cover every camera angle on day one; one well-placed camera per decision point is plenty.
- 03If the AI replaces a judgement call, its alert needs to land where the decision-maker already looks.