Use case / Coffee outlet, India

Know every cup served: edge AI on a coffee outlet's existing CCTV

A coffee outlet in India trades about twenty hours a day. Its only record of a sale was the till. Cups handed over without a bill left no trace, and empty glasses on the counters were noticed late.

ClientCoffee outlet, India, open about 20 hours a day
Delivered on7 existing CCTV cameras, 1 computer at the outlet
Models3 YOLOv8 models: cup, empty glass, person
Our roleDesign, model training, installation and tuning
The problem

A cup that went out unbilled showed up nowhere

Each cup is a small sale made in seconds by a team that both prepares and serves. Once made, a cup costs almost nothing more, so every unbilled cup is revenue lost at close to full margin. And customers judge hygiene by what is on the counter in front of them.

Nobody can watch a counter for twenty hours.

What we chose, and why
  • Use the cameras already on the wall. No new hardware at the counter, no change to how staff work.
  • Process video at the outlet. Only counts and reconciliation figures are stored centrally.
  • Count the serve, not the cup. The owner's question was where cups leave without a bill, so the system counts handovers and compares them with the till.
  • Train on the outlet's own footage, so the models learn its cups, glasses, light and angles.
How it works

Eight requirements, all in service

Existing cameras, one edge computer, and reconciliation with the POS.
RequirementWhat was delivered
Count cups servedCup transfer counting on the serving-counter camera
Count only real servesA cup counts only when it crosses from the staff side to the customer side
Spot empty glassesDetection across five counter cameras, every few seconds
Tell someoneOne email if a glass sits uncleared for 5 minutes
Staff versus customersPeople counted inside and outside the service area
Compare with the tillPOS figures sampled every 5 minutes, side by side with the AI count
Daily reportOn screen, exportable to PDF and Excel
Run unattendedStarts by itself and restarts any camera or process that fails
Equipment

What runs at the outlet

ItemSpecification
ComputerDesktop, Windows 11 Pro, 32 GB memory
Graphics cardNVIDIA RTX 3050, 6 GB
CamerasSeven existing CCTV cameras: five counters, one serving point, one service area
NetworkCameras on their own connection, separate from the internet line
DatabaseHosted PostgreSQL for counts, POS samples and reports
Counting a serve

Direction decides what counts

An invisible boundary separates the staff side from the customer side. A cup handed across it counts once. A cup moving back is not counted, and one that lingers on the line is counted once. The running total survives a power cut and resets at the start of each trading day.

Only a handover from staff to customer counts as a serve.
Getting it right

The first method counted almost double

It counted what it saw in a zone, not what was handed over. We replaced it with tracking: each cup is followed and counted once, on direction. Smaller fixes followed the same habit: find the cause on the outlet's footage, fix it, check again.

Found at the outletWhat we did
Cups missed behind a handTracking retuned to follow a cup through brief blockages
Glasses took up to a minute to showDetection cycle cut to a few seconds
Service boundary did not match the counterRedrawn on the live camera view
Two machines writing countsOnly one copy can now run per machine
What changed

Questions the owner can now answer

  • Cups served per hour and per day, independent of the till
  • The gap between cups served and cups billed, for any period
  • Empty-glass incidents per counter, with time and photo
  • How long an uncleared glass sat before someone attended to it
  • Counter staffing through the day

None of these were available before installation. Timestamped counts and photos also replace recollection when something has to be discussed with staff.

Where the return comes from

Mostly from cups that now get billed

Unbilled cups multiplied by the selling price, over the trading year. Smaller returns come from labour no longer spent watching, glasses cleared sooner, and evidence in place of opinion. The one-off cost is the computer and the installation.

Four sources of return, largest first.
Illustrative only: 1,000 cups a day, 365 days, assumed unbilled shares.
Sizing it for your outlet

Your gap, your price

StepWhere the figure comes from
Cups served per dayThe system's own daily count
Cups billed per dayPOS dine-in total for the same day
Unbilled cups per dayThe difference between the two
Revenue recovered per yearUnbilled cups x price x trading days

The chart is arithmetic, not this outlet's result. A chain in the Gulf would use its own POS and camera makes; the approach stays the same.

FAQ

Questions we are asked

Did the outlet need new cameras?

No. The system reads the seven CCTV cameras that were already installed: five on the counters, one on the serving point and one on the service area.

Does video leave the outlet?

No. The three detection models run on one computer in the shop. Only counts and reconciliation figures are stored centrally.

How is a serve counted?

Each cup is tracked and counted once, only when it crosses an invisible line from the staff side to the customer side of the counter.

How is the return calculated?

Cups served minus cups billed gives the unbilled cups per day. Multiply by the selling price and the number of trading days.

Have a problem you want measured?

Tell us what you want to count, check or protect. An engineer will reply, not a sales script.