Lesson 14 — every production AI system has twelve components. Vendors sell one of them. Most projects fail in the other eleven. This is the map the architect uses to scope a build, defend a tier, and earn the mandate. ## The LLM is component number three Open any AI marketing deck and the language is some variation of “powered by GPT-4o” or “built on Llama-3.” The model takes the whole frame. That is component 3 of 12. The other 11 are where 80–90% of project cost, complexity, and compliance risk lives — UI, backend, orchestration, RAG, tools, data, auth, workflows, observability, governance, and infrastructure. They are not optional, they are not negotiable, and they are not what vendors want to talk about. This lesson visualises the reference architecture. Hover or click any card to read what it owns, what tooling it defaults to in a Maltese Tier-2 build, and where the real risks live. Filter by triage tier to watch which components shift to sovereign alternatives when the data classification changes — the same diagram, different infrastructure under it. > Selling “the AI model” is selling 10% of the problem. The architect’s first job is to make the other 90% legible — to the client, to the team, and to whoever signs off on the production-readiness checklist. ## The twelve components Click a card to spotlight it. The size of each card reflects its share of a typical regulated Tier-2 build — UI and backend each carry around 8–10%; RAG and governance each carry 10–12%; the LLM itself is roughly 12%. The four tier buttons above the grid switch the architecture between public-cloud, EU-residency, sovereign-on-MAGNET, and air-gapped configurations. components.json missing ## Reading the diagram through the triage lens The four tier buttons map directly onto the [three-axis triage](/concepts/three-axis-triage/) the architect uses to classify any new engagement. The most useful exercise is to flip between tiers and watch which boxes change colour. - **L0/L1 + N0/N1 (Public / Internal).** Standard Azure West Europe deployment. Azure OpenAI for the LLM, Azure AI Search for RAG. Nothing shifts. This is the default scope of a Tier-1 readiness assessment. - **L2 + N0/N1 (Confidential).** Still Azure West Europe, but data handling tightens — Key Vault Managed HSM, tokenisation at ingest, per-tenant key envelopes. Database, auth/security, and governance all shift to stricter configurations. Tooling stays familiar; controls harden. - **L2/L3 + N2/N3 (Restricted, sovereign).** Now the architecture meaningfully changes. The LLM moves off Azure OpenAI to a self-hosted Llama or Mistral on Azure Local on MAGNET. RAG moves to pgvector on PostgreSQL. Infrastructure stops being public cloud. Governance gets a fuller compliance surface (FRIA, model cards, human-in-loop). This is the tier where most generalist Maltese consultancies have nothing useful to say. - **L4 + any (Top Secret, air-gapped).** Fully on-prem, no third-party APIs, on-prem identity and key management. Tools and integrations close down to a vetted internal surface. The architecture is the same shape — twelve components, same responsibilities — but every line crosses a sovereign boundary. ## The architect’s positioning Three claims come out of this diagram that are worth keeping in the front of every client conversation: - **Repeatability.** Every build is the same twelve components. The work is in mapping the client’s domain onto the diagram and choosing the per-component tooling. That is not heroics; it is a methodology. - **Defensibility.** The diagram makes the compliance surface (component 11) and the sovereignty boundary (component 12) explicit. Regulators recognise the shape. Insurers recognise it. Procurement recognises it. “Powered by GPT-4o” doesn’t survive an MDIA or IDPC conversation; “twelve components, here is each one, here is how it maps to LN 226” does. - **Fast scoping.** A Tier-1 readiness assessment walks the client through their twelve components once and lists what is missing. That conversation is repeatable, time-boxed, and produces a defensible scope for the Tier-2 build. The diagram is the contract surface. ## What this lesson doesn’t say - **It doesn’t say the cost shares are measured.** They are architect estimates for a typical Maltese Tier-2 regulated build. Different engagements have different distributions — a chat-only product weights component 1 (UI) higher; a heavy-integration agent weights components 4 and 6 higher. The point of the weighting is to argue against component 3 owning the conversation, not to claim three significant figures of precision. - **It doesn’t say twelve is the magic number.** Pick any taxonomy with the same shape — Azure Foundry has its own breakdown, AWS has another. The defensible claim is that there is a small number of components, each with its own ownership, each with its own failure mode, and the architect needs a coherent map of all of them. Twelve is our particular map; the principle is what matters. - **It doesn’t say the per-component tooling is locked.** Azure shows up as the default because most current MITA-aligned engagements default there. Sovereign deployments shift the stack; open-source-first engagements shift it again. The point is that there is *a* tooling decision per component, and the architect owns making it explicit. ## Opening the Production Architecture track This is the first lesson of Layer 12 — the synthesis layer that ties everything before it together. The next three lessons drill into the components that need a live demo, not a diagram: - **Lesson 15 — Observability (component 10).** The trace that explains a decision. Reuses the 15 lesson-09 traces with TTFT/TPOT, tokens, cost, and an Article 12 / Article 86 compliance overlay. - **Lesson 16 — Cost (cuts across all twelve).** Per-task cost calculator. Real $ figures from lessons 09–13, sliders for retry / fallback / volume, side-by-side gpt-4o-mini vs gpt-4o vs self-hosted-Llama. - **Lesson 17 — Security (component 8, surfaces 5 and 6).** Indirect prompt injection through a poisoned chunk in the same Maltese corpus. Watch the model get owned; toggle the output validator on; watch the guard catch it. Together with the thirteen lessons before them, the Production Architecture track closes Phase 1 of the practitioner path. Layer 1 taught how the model works; Layer 9 taught how to ground it; Layer 11 taught how to measure it; Layer 12 teaches how to ship it.