In June 2026, I wrote a design document for what I called an Autonomous Semantic Vector Graphics Engine : a locally hosted multimodal agent where an LLM acts as an architect, decomposes a prompt into named components, draws them in a see-think-act loop, and hands you back clean, semantic SVG.

It was, in the reviewer's words, "a well-organised brainstorm, not a buildable spec."

Four months later, there is a working engine. It is not the thing in the document. It is a constraint-solving vector compiler with a text front door I now use as a smoke test, and a sketch front door I trust. This is the story of how it turned around - including the parts where I was wrong and the parts I still haven't finished.

The Pitch

The document opened with four problems, and they were real:

Problem 1 - Context Window Bloat.…