The image is half the work: production is the finish line
General-purpose image models can already produce beautiful pictures of jewelry. But a picture cannot be sent to production: it carries no metal or stone specification, no dimensions, no notion of manufacturability. Floating settings and uncastable structures look fine on screen and only reveal themselves at the workshop. Our research begins where the image ends.
The architecture
At the core of mulapis is an in-house AI orchestration engine. A fine-tuned orchestrator model dispatches each task to the frontier generation model best suited to it. Every result then passes through a verification stage that checks it against manufacturing specifications: structure, proportion, setting. Designs that fail are never shown to the user; they are regenerated until they pass.
A specification that runs through everything
Running through the whole pipeline is a single design specification: metal, stones, dimensions and quotation, generated and confirmed alongside the design itself. On mulapis, a design is never just an image. From the moment it exists, it is an executable specification.
From design to 3D production files
Once a design is confirmed, the engine exports 3D production files for the production engineer. From conversation to file, nothing needs to be remodelled and nothing is lost in translation.
Why we built it this way
Generation models keep changing; the strongest model today may not lead next year. So we place no bet on any single model: orchestration chooses, verification guards, the specification carries. As models improve, mulapis improves with them, while the road from design to production remains ours.