The single beast, resident in memory
Gemma 4 26B A4B loaded once in LM Studio, serving an OpenAI-compatible API at 127.0.0.1:1234. The developer log streams a design document being generated live, while Activity Monitor (right) shows the memory pressure of running the whole stack locally.
gemma-4-26b-a4b-it-mlx
The instruction-tuned, active-parameter Gemma build — MLX-quantized for Apple silicon. The logs show the model emitting non-functional-requirement and SLA tables for a Copilot design.
A design document, mid-build
The Aether chat UI generating a Zero Trust Network Access solution design — searching the knowledge base, performing live web search, and rendering architecture diagrams. Quick-example buttons line the sidebar.
Computed confidence, in the open
The expandable metadata panel: routed to the Microsoft DWP Technology Architect, 90% confidence, live web search performed, and the exact Microsoft sources cited. Trust you can read off the screen.
The Copilot readiness design doc
A finished “Microsoft 365 Copilot Readiness & Deployment — Solution Design Document,” complete with a Document Control table, owners, reviewers and classification — produced from the universal template, not a placeholder in sight.
Senior-stakeholder framing
The Executive Summary section: business problem, high-level solution, expected outcomes, licensing, and success criteria / KPIs — the structure the design-doc mandate enforces on every generation.
The universal design template
The Design Document template.md that the orchestrator injects into the prompt — sections for Document Control, Executive Summary, current-vs-target state, risk, NFRs, compliance mapping and more. Agents must fill every section or write “Not Applicable.”
One namespace per domain
The kb/ directory: a walled folder for each agent — aws, azure, cloud, core_networking, dwp, enterprise, euv, gcp, microsoft_dwp, network and the network specialists — plus shared design_template and design_docs. Knowledge isolation as anti-hallucination architecture.
The Law of the Jungle
The behavioural Rules.md the assistant follows when changing the code: think before acting, simplicity first, surgical changes, goal-driven and verified. A few rules everyone keeps — because the alternative is chaos.