Tessera Viz
PX-02
Supply chain and logistics
See every shipment, every route, every risk.
This is the one discipline of the five that starts on the map rather than the graph. The network still matters — it is where the single points of failure are — but the question usually begins as a geography.
01
03
Your routes on a map — and the risks hiding in them.
01
Map-first
Weighted flows between places, width by volume, with edge bundling that groups parallel routes into visible corridors. Bundling runs in projected world space, so it stays correct at every zoom instead of re-bundling on each pan.
02
Tracked movement
Entities with interpolated positions between samples, and trails behind them. Interpolation is strictly between known samples — we do not invent a position beyond your data, because on a map an extrapolated position is a false claim about where something was.
03
Clustering by zoom
A dense port is one marker labelled with a count rather than four thousand overlapping markers.
04
Schedule against actual
Both on one timeline. Brush a window and the map shows only what moved in it.
05
Single points of failure
The dependency graph, with betweenness over it. The supplier everything routes through is the one the algorithm finds, not the one someone remembered.
06
Unplaced data
Entities with no coordinates go into a side panel rather than at zero-zero in the Atlantic. A missing location is a fact about the data, and it should look like one.
02
And with an agent
03
Ask in words. Get the flow back.
01
What it changes here
“Which routes feed this plant, and what happens if the second tier goes down” comes back as flows on a map with the corridors bundled — a picture of the disruption rather than a list of the parts of it.
02
A document, not a picture
The model returns one JSON document — the graph, the layout, the algorithm, the theme — and the library draws it in the page. Nothing renders server-side, so there is no image to regenerate when the next question changes the filter.
03
Checked before it draws
The document is validated first. A field the model invented comes back as a list of issues rather than as a blank panel or a thrown error, which is the difference between an agent that can be trusted in front of an analyst and one that cannot.
04
It can arrive as it is written
The stream reader takes the document while the model is still producing it, so a large graph starts appearing at the first complete object instead of after the last token.
05
MCP carries it; this draws it
Expose that document as the return value of an MCP tool and an agent draws straight into the conversation. The library does not implement MCP, any more than an image decoder implements HTTP — it is the renderer at the end of whichever transport you already run.
06
Why this is possible at all
Because the drawing runs in the browser and takes its instructions as data. A library that renders server-side hands an agent a picture; this one hands it a surface that the next question can change.