World-builder

A generation system that turns a single learning goal into a complete training path: a knowledge graph of the concepts inside it, the scenarios that exercise them, the objectives and stages each scenario breaks into, the simulated app the person works inside, and the rubrics that grade everything.

Hand-authoring every curriculum for every user does not scale, and neither does asking a designer to draw the knowledge graph by hand for every new domain. The platform needs to take a short description of what someone wants to learn and generate the world around it, including the way the platform will know whether they actually understood.

A FastAPI service backed by Redis Queue workers on AWS Fargate. Several engines run inside it. A focus-map engine decomposes the target skill into a knowledge graph of concepts and prerequisite edges. A foundational-scenario engine runs a parallel multi-agent research workflow over an unfamiliar domain and emits scenario seeds for the rest of the pipeline. Downstream engines then walk that graph and those seeds top-down (focus to scenario to objectives to stages), persisting each entity to Supabase before the next runs, using Gemini 3 Pro and Claude Sonnet 4.6. Understanding rubrics are auto-generated per entity from a single Pydantic-validated prompt template so the entire rubric set can be rebuilt with one call when the template improves.

Co-contributor on the world-builder team. Contributed to the focus generation engine and the auto-generated understanding rubric path that every other engine depends on. Co-built the focus-map and foundational-scenario engines. Across these engines, focused on raising the quality of what gets generated so the rest of the platform can rely on it.