LLM architecture and orchestration
Structuring agents, context and tools into a system that has to stay predictable.
The hard part of building with language models is not the API call: it is designing a system whose behaviour you can predict, test and explain to the people who depend on it.
In the final build the article body comes from PostgreSQL as semantic HTML, server-rendered so crawlers read the full text without depending on the 3D layer.
The article view drops the constellation effects and prioritises reading: 60 to 72 character measure, 1.78 line height, contrast above 4.5:1. The origin star stays noted in the footer so the reader keeps their place on the timeline.
On close, the camera returns to the exact point it left from, with no constellation reload.