Vision AI · sensor fusion
A cooler that checks its own work before it charges anyone.
Vision AI is the AI smart cooler technology inside Kava Cooler and Kava Box. Edge cameras see what moved, per-shelf load cells confirm what left, and an on-device processor reconciles the two before the receipt lands, within 10 seconds of the door closing.
- checkout accuracy
- 99.4%checkout accuracy
- load-cell resolution per shelf
- ±2 gload-cell resolution per shelf
- shrink vs. honor-system markets
- −62%shrink vs. honor-system markets
How AI smart cooler technology turns a door-open into a receipt
Tap and open
A card, phone, app QR or badge is authorized and the electromagnetic lock releases.
Watch the products
Four edge cameras (six on C2-XL) follow items as they leave and return to the shelves. They track products, not the person holding them.
Weigh the change
Per-shelf load cells register each change at ±2 g resolution, so a returned item and a taken item never look the same.
Reconcile
The on-device processor fuses both signals into one basket. Where they agree, the basket is final. Where they disagree, the event is flagged instead of guessed.
Receipt
The charge settles for exactly what left the cooler, and the receipt is sent within 10 seconds of the door closing.
Where one sensor fails, the other usually doesn't
| Situation | Cameras only | Weight only | Vision + weight |
|---|---|---|---|
| Two look-alike cans, different flavors | Can miss when labels face away | Can't tell them apart | Weight narrows the shelf, vision picks the flavor |
| Sandwich and salad of similar weight | Tells them apart by shape | Ambiguous | Resolved by vision, confirmed by weight |
| Item put back on the wrong shelf | Sees the return | Sees a loss and a gain on two shelves | No charge; item location updated |
| A hand blocks a camera | Blind for that moment | Still registers the lift | Weight carries the event, vision confirms on exit |
| A restocker loads 40 items | Counts what it can see | Counts total mass | Item-level count, reconciled to the pick list |
Kava Cooler C2 uses 4 edge cameras and C2-XL uses 6, both with per-shelf load cells. For a deeper comparison of sensing methods, including RFID, read computer vision vs. weight sensors vs. RFID.
It looks at products. It doesn't look for people.
Recognition runs on the cooler's own neural processor. The models are trained to find cans, bottles, sandwiches and fruit, and the processing happens inside the cabinet rather than in a data center. No images of faces are stored or sent.
What leaves the device is the result: which items moved, from which shelf, and what the basket was. That's enough to charge correctly, reconcile a restock and spot loss patterns. It isn't enough to identify anyone, and it was designed that way on purpose.
- Recognition runs locally on every Kava Cooler.
- No facial recognition and no stored images of faces.
- Analytics are aggregated across sessions, never built into shopper profiles.
What Vision AI does when nobody is shopping
Restock reconciliation
When a restocker opens the door, the cooler counts what went in, item by item, and compares it with the pick list. Short deliveries and wrong items show up while the van is still in the lot.
Shrink detection
Unpaid removals, held-open doors and repeated basket disputes are grouped by device and hour, so you can see where loss happens instead of guessing from month-end counts.
Planogram drift
Items that migrate to the wrong lane quietly break forecasts and pricing. Vision AI notices the drift and flags the shelf before the next restock.
Dispute evidence
Every basket keeps its shelf-level weight changes and recognition result. A shopper's question gets an answer from data, not from someone's memory of the shelf.
Accuracy you can audit
Where Vision AI runs
Kava Cooler C2
- Cameras
- 4 edge cameras
- Weight
- Per-shelf load cells, ±2 g resolution
- Capacity
- 5 adjustable shelves, up to 240 items
Kava Cooler C2-XL
- Cameras
- 6 edge cameras
- Capacity
- Up to 480 items
Kava Box B160
- Sensing
- Ceiling cameras plus shelf weight sensing
- Assortment
- Up to 700 SKUs across ambient, chilled and frozen zones
Kava Kiosk K1
- Recognition
- Barcode scanner plus vision for unpackaged items such as fruit and bakery
Processing
- Compute
- On-device neural processor; recognition runs locally
- Offline
- 72-hour transaction buffering
Ask any checkout system what it does when it isn't sure.
Every sensor is sometimes wrong. The question that matters is what happens next. Does the system guess and charge, or flag and verify? Overcharging a nurse at 3 a.m. costs more goodwill than the sandwich is worth, and an unbilled item costs less than a dispute. Before you compare accuracy numbers, compare fallbacks. The 0.6% is where trust is won or lost.
Vision AI questions
Does the cooler use facial recognition?
No. Recognition runs on the device and is trained on products, not people. No images of faces are stored or sent. Our security page covers how data is handled end to end.
How is 99.4% accuracy measured?
Basket results are audited against restock counts across Kava Cooler deployments over the trailing 12 months. A restock is a physical count, so it catches errors in both directions: items charged but not taken, and items taken but not charged.
What happens when the cameras and the scales disagree?
The event is flagged rather than charged on a guess. It's resolved from the shelf-level weight record and the next restock count, and the result feeds back into the device's recognition.
Does it work in an open-shelf micro-market?
Kava Kiosk K1 brings vision to open shelving: it recognizes unpackaged items like fruit and bakery and works with existing micro-market fixtures. Moving to Kava Coolers with locking doors is what drives the 62% shrink reduction.
Which plan includes Vision AI?
Checkout recognition is built into the hardware. Vision AI analytics and shrink detection are part of Autopilot at $69 per device per month. See pricing.
Where to go next
Put one cooler where the honor shelf is.
A smart-cooler program of up to 10 units typically goes live 10 business days after signing. Then you can see your shrink number instead of estimating it.