ARCHEON BUILDER

A local-first graph editor that lets you build, annotate, and AI-enrich structured diagrams of complex systems — from brain circuits to EU governance — with typed ports, ontological constraints, and literature-backed evidence on every edge.

Archeon Builder is a tool for people who find themselves drawing the same kind of diagram over and over: boxes connected by arrows representing some complex system. The problem with those diagrams is that they lose information fast. Which end of the arrow is the input? What does this connection actually mean? Is there any evidence for it? Archeon Builder is an attempt to fix that by making the graph itself a first-class data structure — typed, constrained, annotated, and evidence-aware.

The central concept is the archeograph: a graph where edges don't just connect nodes, they connect specific typed ports (called connectors) on nodes. So instead of "A → B," you get "A's output port → B's control port." This sounds like a small change but it matters a lot: it lets you define a meta-architecture — a set of rules about which port types can connect to which — so the graph enforces domain semantics rather than just being a free-form sketch. The same mechanism works whether you're modeling an active inference AI agent, the legislative flow of EU institutions, or the 83-region Desikan–Killiany brain atlas. Those three examples ship as pre-loaded demos.

The AI enrichment pipeline is the most technically interesting piece. When you ask the agent to "add connections from cerebellum to cortical areas from the literature," it doesn't just hallucinate edges. It runs a five-stage loop: infer the domain from your node labels, propose candidate edges with search queries, actually hit PubMed and Crossref, critique each proposal against the real results, then emit only the survivors — each tagged with a confidence score between 0 and 1 and linked to specific papers. A toolbar slider lets you filter out low-confidence edges. Manually drawn edges default to confidence 1.0; agent-drawn edges carry whatever the critique step assigned. This is a meaningful distinction: the graph remembers how sure it is about each claim.

The system runs entirely locally via Docker Compose — a React frontend, a FastAPI backend, and PostgreSQL. There's also a 3D viewer (React Three Fiber) that, for graphs with spatial coordinates like the brain atlas, renders nodes at their actual MNI positions with an optional brain mesh overlay. The data model is clean: all graph objects share a polymorphic supertype, layouts are stored separately from structure (so you can have multiple visual arrangements of the same graph), and the full database exports to a single JSON file for sharing. Current limitations are honest: single-user only, no persistent version history, and the 3D viewer slows down past ~200 nodes.