ZenithAI

01The problem

The report comes from pairing hard NASA JPL astronomical data with text generated by a language model. A model will happily write a graceful sentence about something the data never contained — and the customer pays up front and receives a file nobody will correct afterwards. The work was in separating what the model is allowed to invent from what has to come out of a calculation.

02What I built

• Every numeric value is computed application-side from the JPL data; the model receives them as settled facts, not as a sum to work out.
• The model is responsible for the narrative layer alone — the shape and language of the report, never the inputs.
• A React configurator walks the user through step by step and validates the inputs before payment is allowed.
• Full PL/EN bilingualism at the level of content and of the generated report, not just the interface.

03What it does now

• The figures in the report are reproducible — the same input set always yields the same result.
• A purchase ends with the file delivered, with nobody on the seller's side involved.
• Both language versions come out of a single generation pass.