Stock is cash tied up, and in most companies it is decided in Excel. Rungle set out to answer three questions about a multi-tier network: where the stock is, where it should be, when to move it. Behind it, a constraint solver. In front, the problem with this kind of product: constrained optimisation is an engineer's subject, and the people who know the real constraints are planners. As long as the rule has to be written by a solver specialist, the product never leaves the lab.

The rule, and its translation underneath

The rule editor: the sentence “An harbour cannot be more than 110% full”, underneath a rephrasing, its MiniZinc translation and the linked parameters, Boats and Harbours
The rule editor, on the working board. The example is a harbour and its boats; the real case was warehouses and pallets.
A harbour cannot be more than 110% full.
Sum(harbour.boats.size) < harbour.capacity
Parameters involved: Boats, size and home harbour · Harbours, capacity and boats.

The user writes the rule the way they would say it. Just underneath, the interface shows what the solver understood, in its own language, with the parameters the sentence calls on. The user keeps control of the translation: if it is wrong, they see it before running anything. The language model is there to formalise the rule and to explain the recommendation. The decision itself stays computed and verifiable.

The model before the screens

Anything that can be written in a rule has to exist in the model. Four notions are enough.

EntitiesWhat gets moved or transformed. A product, an order, a batch.
ResourcesWhat is used to do it. A machine, a lorry, an operator.
ActivitiesThe processes that act on entities. Assemble, transport, store.
EventsThe state changes that start or interrupt an activity.
The resource editor: Boats and Harbour with their parameters, the form for adding a parameter with its type, domain, value to optimise and rules
Resources and their parameters. Each parameter has a type, a domain, and can be the value to optimise.
The four domain definitions and the SQL schema: Entity, Flow, Ressource, Allocation, Trigger, Activity, Event
The four definitions and the data schema, laid out next to the screens on the same board.

The running application

A capture of the application: scenario cards for France, 500 scenarios over 12 weeks, with margin, loss, days of stock, stock-outs
The application running locally: scenario cards, five hundred scenarios over twelve weeks, with margin, loss, days of stock and stock-out rate. Sample data.

What the field taught

“SKU” does not mean the same thing to different teams within the same company. A brand selling on several marketplaces cannot arbitrate its stock between them, because the policies and promotion windows differ. Forecasts are assumptions, and they are better treated as such, signed, than as data.

The stopping criterion

A pilot customer signed by the end of January 2026, or we stop.

The criterion was set before starting, by the two of us, so that we would not have to decide it at the moment it would hurt. At the end of January, no pilot customer. Rungle stopped in early 2026.