Five production systemsone shop floor
Goods come in through the first one, get counted and routed by the second, get manufactured and booked by the third, get verified on the way out by the fourth, and the fifth watches the other four and says something when they quietly go wrong.
Every demo below is interactive and runs on invented data. The company is not named. Four of them have a switch that lets you break the process on purpose and watch the system refuse to be fooled. The ledger lets you work a batch from rack to receipt.
Use AI to see. Use code to decide.
A vision model is very good at reading a small printed code off a box at an angle, and very bad at being the last word on whether you should ship. So it never is. Every hard lesson in this portfolio points the same way: when a model reads the discriminating text correctly and then reasons past it, the answer is a deterministic check, never a better prompt.Dock to storefront intake
Photograph a pallet. Get a lot-traced listing and a booked receipt, without anyone typing a part number.
By my own count, three people down to one on a phone, about an hour of combined work per pallet, estimated. On the reference intake a human re-read every label and matched the machine lot for lot across 180,000 units. break itSwap one photograph for a blank page and watch approval refuse to proceed. 02Daily inventory sync
The production plan, correct three times a day, with nobody exporting anything.
About a thousand reorder rows a run. The interesting part is the rule I deleted: it only ever fired on a row a human had deliberately changed. break itTurn the legacy rule back on and watch your own reassignment get quietly reverted. 03Rack to receipt ledger
A stock rack that was in no system at all, turned into a full production lifecycle with printed travelers and labels.
While designing it, I found 101 units of demand on the wrong work station. Nobody had done anything wrong. Nothing had been watching. try itCut a batch, print a label, scan the barcode with your phone. 04Shipping verification kiosk
Every outbound order photographed, counted without the expected quantity, and checked against the order by code rather than by confidence alone.
First-try pass of 78.5 to 87.5 percent by month and station, at about $0.15 of model spend per order at list rates. Built to be time neutral for the operator on purpose. break itChoose the variant trap and ship the wrong variant, watch the check stop it, then turn the check off and watch the name match pass it. 05Operations watcher
It watches for the failure that leaves the dashboard green: the system said something was wrong, and nobody stopped.
Scans every five minutes in working hours for known failure classes. Its dark-station rule was written after a machine sat powered off for a weekend. break itSimulate a station going dark and watch a normal dashboard stay green.