They rot.
Drawn by hand, accurate on the day of the review, wrong after the next change window. Nobody trusts a diagram without checking its date.
The copilot for network and infrastructure architects
Ask for the diagram you need. Get the one your inventory can prove.
Describe the view in plain language. The AI turns your sentence into a specification you can read and correct. A deterministic graph engine does the rest — from your source of truth, and nothing else.
The AI decides what to ask for. The graph decides what exists.
The problem
Three reasons network diagrams fail
The full map is unreadable.
Hundreds of devices, thousands of links. Every audience needs a different cut, and each cut costs someone a day.
Generative AI draws plausible networks.
Ask a chatbot for a diagram and you get links that look right. In network architecture, a link that looks right and does not exist is an incident waiting to happen.
How it works
From a sentence to a diagram, without losing the truth
- AI
Ask
“Show L3 connectivity between Paris and Singapore, including firewalls.” Then refine as in a conversation: add a site, hide a tier, simplify for a review board.
- AI → spec
Interpret
The AI produces a structured specification: sites, layers, device types, level of detail. You see it, you can edit every field, and what it did not understand is reported — never guessed.
- Graph engine
Query
A deterministic engine executes that specification on the graph of your inventory: filtering, path finding between sites, grouping. Same request, same result.
- Layout · export
Deliver
Automatic layout by site and network tier, three levels of abstraction from detailed to executive, export to draw.io and PNG.
The AI decides what to ask for. The graph decides what exists.
Trust
Every link on screen has a receipt
Network teams do not need another tool that guesses. The product is built around one rule: nothing is drawn unless the inventory says so.
The AI never sees your topology
It receives a vocabulary — site names, device types, what the engine can do — and your sentence. No device, no IP address, no link is sent to it. The product also runs with a rule-based interpreter and no external AI call at all.
Source fact or view decision, always labelled
A link recorded in the inventory is a source fact. Twelve access switches shown as one node are a view decision: drawn differently, and listing the devices it stands for.
No invented shortcut
Hide the distribution layer and the diagram stays disconnected, rather than showing a core-to-access link that does not exist. When Paris reaches Singapore through Frankfurt or London, the transit is shown even if you never named those sites.
Checked on every request
Each view passes an integrity check before it reaches the screen: every connector maps to relations of the inventory, with the same type and the same endpoints. Otherwise the request fails.
248 → 20nodes on screen when asking for the backbone only
2transit sites revealed between Paris and Singapore, which the request never named
0links drawn without a source relation, by construction and by test
Figures from the demo dataset (synthetic network).
Use cases
Built for the moments when a diagram decides something
Architecture review boards
One network, the right cut for each audience: detailed for engineers, aggregated for the review board, executive for the steering committee.
Change and impact preparation
What sits between these two sites? Which firewalls are on the path? Ask, check the interpretation, attach the view to the change.
Security reviews
Firewalls, security zones and Internet attachment, grouped by zone. What is hidden is listed, so nobody mistakes a filter for an absence.
Audits and onboarding
Saved views run on today's inventory and say what changed since they were saved. Newcomers explore the network by asking questions.
Collaboration
Views that do not rot, deliverables that keep their proof
Saved as a question, not as a drawing
A saved view is its specification and the conversation behind it. Re-open it in six months: it runs on the current inventory and reports the drift — devices and links added or removed.
Shared with a link
Colleagues open the same view and refine it in their own conversation. The original only changes when someone decides to overwrite it.
Exported to draw.io
Cisco stencils, site containers, interface and IP labels. Each shape carries its provenance in its data, so the hand-finished deliverable stays traceable.
Integration
Your source of truth stays where it is
The product reads your inventory through a single adapter and keeps a canonical graph: sites, devices, interfaces, VLANs, subnets, security zones, applications. The large language model is interchangeable too.
Connectors planned with design partners
- NetBox
- Nautobot
- ServiceNow CMDB
- LeanIX
FAQ
Questions architects ask first
Does the AI see my network?
No. The language model receives your sentence and a vocabulary: site names, device types, relation types, what the engine can do. Devices, addresses and links stay in the graph engine. You can also run the product without any external AI call, using the rule-based interpreter.
Which AI model is used?
The interpreter is a replaceable component. The current version includes an adapter for Claude (Anthropic) whose answers are constrained to a strict schema, and a rule-based interpreter that works offline. Other providers can be plugged in.
What if the AI misunderstands my request?
You see it immediately. Every answer comes with “I interpreted your request as…”: sites, layers, included and excluded device types, level of detail. Each field can be changed by hand, and manual changes bypass the AI entirely. Words that were not understood are listed instead of being guessed.
Which inventories are supported?
Today the demo runs on a synthetic dataset. The inventory is read through one adapter, designed for sources such as NetBox, Nautobot, ServiceNow CMDB or LeanIX. The first connectors will be built with design partners.
Can I keep my draw.io deliverables?
Yes. Any view can be exported as a draw.io file or a PNG. The draw.io export is a dated, frozen copy that keeps the provenance of each element and a link back to the live view.
Is it production-ready?
No, and we prefer to say so. It is a proof of concept: single user, no access control, in-memory graph. A pilot starts from a read-only extract of your inventory.
Bring three questions about your network.
We run the demo live on the synthetic network, then look at your inventory and at what a pilot could be.
Request a demo