Skip to content
focoit.com
Focoit home
AI · ArchitectureNovember 18, 2026

AI architecture: design the system around the model

The short answer first; the detail is below it.

Short answer: the model is one box on the diagram. Most of the work, the cost and the risk in a production AI system sit in the architecture around it — grounding, orchestration, evaluation, guardrails, identity and the operating model. I design that system on Azure, then prove the risky parts in code before the design is committed.

The hard part of AI is rarely the model; it is the operating model around it. A production LLM system needs retrieval, orchestration, evaluation and guardrails around the model — and those are architecture decisions, not model settings. That is the part I work on.

Who it is for

Teams with an AI use case and a pilot that has not shipped. Non-technical founders deciding whether to build with AI at all. Engineering leads who have been sold a model and need an architecture around it.

It is also for teams in healthcare and other regulated industries, where privacy, security and model risk are part of the design brief. I bring the Azure and AI architecture; your compliance owners bring the compliance decisions. If you already have a working prototype and the question is what to keep, that is a good starting point too.

What you get

  • A reference architecture. Agentic or traditional RAG, vector search, and multi-agent orchestration with Microsoft Foundry IQ and Semantic Kernel — chosen for the problem, not for fashion.
  • An honest retrieval comparison. Agentic against traditional RAG, evaluated on token efficiency and response quality, instead of assuming the more complex option is better.
  • Evaluation and guardrails. Responsible AI and Zero Trust built in, with evaluation and observability across the components — so you can see what the system did and why.
  • A governance baseline. The Microsoft Cloud Adoption Framework as the starting point for security, governance and operations.
  • A build, buy or configure call. Whether the system should be built, configured on an existing platform, or not built yet.
  • Build-to-run. CI/CD, environment strategy, observability and cost transparency, so what ships transitions cleanly into long-term operation.

On Azure that often means Azure OpenAI or AI Foundry for the model, a vector search index for retrieval, and Semantic Kernel for orchestration — with Zero Trust access and the Cloud Adoption Framework around all of it. The pieces are chosen to fit the problem, not copied from a reference diagram.

The architecture is written so you can hand it to any team — including the one you already have.

How it works

  1. Strategy Call. Forty minutes on the goal, the data and the constraints. No deck required.
  2. Architecture and assessment. A scoped Azure architecture for the AI system, a realistic cost view, and a build-vs-buy recommendation. The assessment covers the data you can use, where it lives, who can access it, and what the model is allowed to see — because data and permissions decide what the system can safely answer.
  3. Prove the risky part. I prototype the pieces most likely to fail — in code, configuration, pipelines and prompts — before the design is signed off.
  4. Build or hand over. I architect and deliver on Azure with security and governance from the start, or hand the design to your team to build.

What it is not

  • Not an AI strategy deck. You get something you can act on, not a slide about transformation.
  • Not a vendor pitch. The honest answer is often to configure an existing platform, buy a component, or wait — and you will hear it.
  • Not a promise that AI is the answer. A system that should not exist is still a bad system, however good the model.
  • Not compliance expertise. I am compliance-aware, not a compliance auditor. I design within Azure's capabilities, alongside your compliance owners, and I don't give legal advice.
  • Not a demo that stays a demo. If it cannot reach production with the security, cost and operations around it, the architecture is not finished.
  • Not a model-training project. The work is the system around the model: retrieval, orchestration, evaluation and operations.

If the system should not be built, that is the recommendation — with the reasoning written down.

Next step

Bring the use case, the data you can actually use, and the constraints you already know about. Forty minutes is enough to say whether the system you have in mind is the right one — and what to prove first either way.

Book a Strategy Call