Knowledge Graph System knowledge visualisation

Knowledge Graph is an information-design piece, not a product. It draws a connected body of my own work — the parlance platform, the lexicon design system, the the-pace sites and their specs — from a knowledge base of roughly 1,600 notes and 1.5k source files, and lays it out as a single force-directed map. It exists to answer one question that folders and links can never answer at a glance: what is connected to what? The raw graph already existed, auto-extracted and auto-clustered by tooling. My contribution was the visual grammar that turns an accurate-but-noisy data structure into a picture a person can actually read.

Systems thinking has an output problem. The relationships between projects, systems and decisions are real, but they live in folders, imports and cross-links where nobody can see the shape of them. You can feel that parlance depends on lexicon, or that a spec note anchors three features, but you cannot point at it. Without a picture, the structure stays in one person’s head and leaves when they do. And the naive fix — let a tool draw the graph — trades one illegibility for another: an honest force-directed drawing of ~1,600 connected nodes is a grey hairball, while a pretty drawing usually lies about the data.

• Relationships between projects and systems are invisible in folders and imports
• The structure of a body of work lives only in one person’s head
• An honest drawing of ~1,600 nodes is a hairball; a pretty one usually lies
• Default clustering over-fragments — meaning is there, but scattered and unnameable
• Importance and community are lost in a flat, undifferentiated link list

The raw graph already existed — auto-extracted from ~1,600 notes and source files, then auto-clustered by tooling. So the research wasn’t ‘what is the data’; it was entirely about legibility: what a 1,594-node graph must encode to be read at a glance, what to suppress so it isn’t a hairball, how to recluster an over-fragmented partition into nameable groups, and how a stranger can name the structure in seconds.

I started by auditing the raw material rather than restyling it. The extractor had produced 1,594 nodes and 1,606 edges at 85% direct extraction, but the default partition was over-fragmented, so the first move was a second clustering pass: Louvain at a fixed low resolution, then small fragments merged into the nearest large cluster by source-file path similarity — entirely algorithmic, no hand-placing. That pulled the estate into 37 nameable communities. From there the work was pure encoding. I gave each community a colour from a ten-hue categorical palette, sized every node by its connection count and added a glow on dark so the hubs read without a single label, muted the edges so connection shows without noise, reserved always-on labels for the hubs alone while the rest stay silent until you reach for them, and wrote a legend that names each community by a representative hub node. Even the estate’s honest edges got a home: the 171 nodes with no edges at all are edgeless by extraction, not by choice, so instead of scattering them I gave them their own labelled bucket.

The recurring tension in any knowledge graph is that accuracy and readability pull against each other, and the honest move is usually the less flattering one. Reclustering makes the estate far more legible, but it does not — and should not — erase the truths the data holds. lexicon still surfaces across several communities because it genuinely touches several parts of the estate; 171 notes carry no edges at all because extraction found none. The temptation is to tidy those away for a cleaner picture. I chose instead to surface them honestly — lexicon’s spread reads as reach, the edgeless notes get their own named bucket — so the map stays a truthful instrument rather than a flattering poster.

• Keeping ~1,600 nodes readable without hiding the truth of the data
• Rationing labels to hubs so shape reads first and detail reads on request
• Reclustering for legibility while refusing to erase what the graph honestly shows
• Surfacing edgeless nodes and cross-community reach as insights, not defects to tidy away
• Making a technical, auto-generated artifact feel like a deliberately designed object

The final piece is a dark, glowing, clustered map of the whole estate in one frame: parlance, lexicon and the the-pace sites drawn as 1,594 nodes and 1,606 edges, coloured into 37 communities from a ten-hue categorical palette, sized and glowed by connection count so hubs like POST(), ParlanceClient and the parlance platform node read instantly, edges muted, labels reserved for the hubs, and a legend naming each community by a representative hub node. It is honest by design: lexicon appears across several communities because it truly reaches across the estate, and the 171 edgeless notes sit in their own labelled bucket rather than being hidden. It works as both a navigational instrument — click a hub, its neighbourhood lights up, its name appears — and a portfolio statement about how the pieces fit.

Knowledge Graph is a small but pointed demonstration of systems thinking made visible: an accurate, unreadable data structure turned into a designed picture that communicates the shape of an entire body of work at a glance. Its value is legibility, not new data — nothing in the map isn’t already in the estate’s files and imports. The measurable facts are the corpus itself (1,594 nodes, 1,606 edges, 37 communities, 85% directly extracted), and what the design adds is the ability to read them: to see in one frame that parlance is the estate’s centre of gravity, that lexicon reaches across it, and that a long tail of notes still sits unconnected — each of them a true observation the picture surfaces rather than an artefact it invents.

“

You cannot see a system until someone draws it.