smert.ai — AI Cobots, Robotics Integration & AI Consulting

Why vending needed computer vision

Traditional vending lost ground for a decade — limited SKU range, no personalisation, painful restocking economics. In 2025 a wave of computer-vision-equipped machines reopened the category, and smert.ai's smart vending platform sits in the middle of that shift.

The hardware stack

Every smert.ai smart vending unit ships with:

Two interior cameras with on-device inference
A weight-sensing shelf array
A glass-front touch display
Edge compute powerful enough to run vision models locally — no cloud round-trip per pick

Shrinkage: down 38%

Across 220 deployed machines, shrinkage dropped from an industry-average ~4.1% of revenue to 2.5% in the first six months. The biggest contributors:

Real-time pick verification (camera confirms what left the shelf)
Automatic flagging of door-open anomalies
Tamper detection that pings the operator within seconds

Restocking cost: down 41%

Vision data tells operators exactly what is left in every machine. The route engine rebuilds itself nightly and only sends a truck to machines that actually need it. Average kilometers driven per restock dropped from 38 to 22.

The recommendation layer nobody expected

Once the camera knows what the user looked at — and the touchscreen shows the right context — average ticket lifts. Across the fleet:

Average ticket: +18%
Repeat visits within 7 days: +27%
Successful upsell from snack to combo: 9.3%

Privacy and the camera question

We do not store identifiable images. All vision processing runs on-device, and only structured events (item picked, item put back, anomaly detected) leave the machine. This is a non-negotiable design choice — and the reason corporate and airport clients let us deploy in the first place.