Steven Johnston
← Work
№ 05 · AI Network Design Studio · 2026

ARCHIE

Upload a customer brief, even a photo of the whiteboard, and designer/critic agent pairs argue their way to a cited network design. A build agent then turns it into Containerlab, GNS3, and draw.io artifacts.

Context
Personal project
Role
Sole designer / engineer
Status
In active use
Design · acme-2026 · run #4
LAN approved
12 citations
WAN critic · round 2/5
1 flaw open
QOS designing…
CLOUD queued
WAN · designer ⇄ critic
designer
Dual-homed DMVPN hubs, EIGRP named mode, per-tunnel QoS.
WAN-DG §4.2QoS-DG §2.1unsourced ×1
critic · strictness 4
Citation WAN-DG §4.2 verified ✓, but no failover timer rationale. Revise.
coherence pass: LAN voice CoS ↔ QOS class map · 1 conflict
Fig. 01 · stylised interface preview

The problem

After a customer meeting, turning scribbled requirements into a defensible design doc takes days. The LLM shortcut produces confident designs with invented best practices. If a design decision can't cite a real guideline, you need to know that before the customer asks.

The approach

01

Multimodal intake: typed PDF, Word, or Markdown, pasted notes, or a photo of the whiteboard via a vision LLM. No separate OCR service. An extractor splits the brief into LAN, WAN, EDGE, CLOUD, and QOS sections.

02

An orchestrator derives a shared baseline (routing family, AS numbers, IP plan, QoS classes, security posture), then spawns one designer/critic loop per section, running in parallel. The designer must cite a ChromaDB corpus of design guidelines. Unsourceable decisions are flagged, never fabricated.

03

The critic independently re-queries ChromaDB to verify each citation in code, so the LLM doesn't grade its own homework, then hunts for design flaws. Strictness is user-controlled 1-5. A coherence pass catches cross-section conflicts and forces revision.

04

Approved sections feed a conversational build agent that decides per-turn whether to ask or build, emitting a topology spec, deterministic Cisco IOS / FRR configs, and exports as draw.io XML, Containerlab YAML, and GNS3 skeletons. Push to gear is lab-inventory-gated with a dry-run then confirm flow.

Outcome

Citation checking
In code
the LLM doesn't grade its own homework
LLM providers
5
Anthropic, OpenAI, xAI, DeepSeek, Ollama
Export targets
3
Containerlab, GNS3, draw.io
Design note

Archie is the colleague who reads the whole design guide, argues with you about it, and then builds the lab to prove the point.

Built with

PythonFastAPIChromaDBClaudeReact 19CytoscapeMermaidnetmikoSlackDocker

More tools

06 CLOUDFORGE Chat-to-Azure IaC 07 DOCKERMATE Docker Dashboard + AI Ops 08 WEBEX MIGRATE CUCM → Webex Calling
← Back to tools
Get in touch
stevie.johnston@gmail.com
Glasgow, Scotland
UTC+0 / UTC+1