)
Packing furniture onto pallets, without the Tetris grand master
Woodi grew from a carpenter shop into a factory, and one job never changed: a handful of people who could see in their heads how an order of school furniture ought to sit on a pallet. Fit it onto four pallets instead of five and the shipping bill drops by up to a fifth. We gave that job to an optimiser.
The problem
Every order leaves on pallets, usually in a container on a lorry, and every pallet has to be planned: logically, tightly, and around the awkward shapes furniture insists on having. The catch is the timing. Planning has to happen before production, and before the order is even final, because sales need an accurate shipping cost for the quotation.
So the work sat with a few people who had developed a sixth sense for it. That arrangement holds until you need two of them at once, or one of them is on holiday, or the intuition has to be taught to somebody new. Loose pallets have a cost of their own: lorries hauling air across the countryside, paid for in both money and diesel.
What we built
No AI, no computer farm, no neuro-wizard. A couple of PhDs wrote a traditional optimisation algorithm and put a usable interface on top of it.
The optimiser works from Woodi’s own history. Previous orders act as templates for new ones, so the existing data sets the packing rules. There is no model to train and no model to maintain, which is most of the benefit of machine learning without the usual upkeep, and every new packing solution added to the database improves the next plan. The system is integrated with the ERP, so sales and the packing floor receive the same plan at the same moment and it is archived for the next order that looks like it.
The hard part was not the algorithm
It was getting people and machine to collaborate sensibly. The obvious interface, here is the suggestion and fix what you don’t like, quietly destroys the point of the system. The moment a user starts repacking a pallet by hand, they are doing the whole job manually again and the machine has learned nothing.
So the user does not repack. When they see a pallet that could be better, they build an example pallet model showing a better way, and run the plan again. The model is taken into use immediately, with no slow, brittle retraining loop.
What comes back may not contain that exact pallet. It will usually have rethought the whole order in light of it, which is a job no person would take on for one improvement and a computer does without complaint. That takes some getting used to, and it is the reason the system keeps improving instead of drifting back to manual work. This was several design iterations and a good deal of testing, not a decision made on a whiteboard.
What changed
Pallet plans are ready early enough to price a quotation, without booking one of a few specific people.
Tighter packing, four pallets where there were five, cuts shipping cost by up to 20% and takes lorries off the road.
Knowledge that lived in a few heads now lives in a database that grows every time somebody teaches it something.
The methods here are unremarkable, and that is the point. The repetitive planning went away, the dependency on a few specialists went with it, and nobody had to run an AI project to get there.
Read the original post