Viert System — The atelier's production and sales in one place

Scheduling is a pure function: it takes the full state (projects, resources, calendar, unavailability) and returns the proposed schedule, the project order and the reason behind each choice. It reads neither the database nor the clock, so the same scenario always gives the same result and runs in tests with no infrastructure. The preview runs the engine over a snapshot of the state and writes nothing until the coordinator confirms. The schedule works in wall-clock time, in whole minutes on a 30-minute grid, with no time zone in the core.

Each project is prioritized by slack (available time minus the work remaining), and every event triggers a full recalculation. The catch is that the coordinators also edit by hand: they lock a time slot, lock a person's whole day, move a fitting. Those decisions must survive the recalculation, pull predecessors in ahead of them, and show when delivery misses its deadline, pointing at the step that drives the delay. A circular dependency is rejected on the spot instead of becoming an error screen.

Finishing work goes out to outside suppliers, who work by lead time: each has a drop-off day, a pick-up day, a cap on pieces at once and a weekly ceiling. Without those limits the engine sent as many tasks as it liked to the same supplier. They became a registry rule and enter the allocation just as internal resources' hours do.

The examples in the requirements document itself became scenario tests, with the text report approved as the reference: if behavior changes, the change has to be deliberate. Property-based tests with random scenarios check invariants, and a performance test caps the time for 100 projects, 14 steps and 15 resources. The app runs its integration suite against an embedded Postgres, without Docker.

The catalog came from a spreadsheet, with many records flagged for review. The AI reads each model's photos and flags mismatches with the record, separating what is visible, what is absent, what the photo can't tell, and what is merely likely. Code, not the AI, translates the reading into the catalog's official criteria. Every applied correction keeps the previous value and the author, and can be undone, one by one or in bulk.

The customer talks to pre-sales on one number and moves to sales on another. On the pre-sales number the 24-hour window is open and the notice goes out as free text. On the sales number the customer has never written, and Meta only accepts an approved message template. The template text is editable: changing it creates a new version, which goes back through approval, and until then the last approved version is sent. The handoff stays off until management switches it on.