AI Engineering Studio

Appearance

The Four-Layer Map ​

📝 Context ​

The one mental model the rest of this site hangs on. When a customer describes an "AI problem," your first job is to locate which layer they're actually talking about — because the layer determines who's in the room, what it costs, and how you explain it.

"AI engineering" is not one field. The most useful framing in 2026 — the one enterprise reference architectures and engineer roadmaps both converge on — separates the application layer (where most engineering jobs now sit) from the infrastructure beneath it, the classic ML/data discipline it grew out of, and the architecture & governance concerns that wrap all of it.

🗺️ The Four Layers ​

flowchart TB
  L4["L4 · AI architecture and governance"]
  L1["L1 · LLM apps and agents"]
  L2["L2 · LLMOps and inference infra"]
  L3["L3 · Classic MLOps and data"]
  L4 --> L1
  L1 --> L2
  L2 --> L3
LayerNameWhat lives hereWho owns it
L1LLM apps & agentsRAG, agents, prompt/context engineering, evals, orchestration, MCP/A2A, multi-agentApp / product engineers
L2LLMOps & inference infraServing (vLLM/SGLang), quantization, GPU, vector DBs, gateways, LLM observability, cost/latencyPlatform / infra engineers
L3Classic MLOps & dataTraining pipelines, feature stores, registries, experiment tracking, CI/CD for ML, driftML / data engineers
L4AI architecture & governanceReference architectures, build-vs-buy, guardrails, evaluation strategy, NIST / EU AI Act / ISO 42001Architects, security, legal, exec

Read the arrows as "depends on." An L1 chatbot depends on L2 serving a model fast and cheaply, which depends on L3 disciplines for anything you train yourself, all bounded by L4 decisions about what you're allowed to build and how you prove it's safe. Most new engineering work sits in L1; most cost and risk sit in L2 and L4 — exactly where an SE earns their keep.

🎯 Why This Matters in the Room ​

Most online "AI engineer roadmaps" collapse L1–L4 into a single 12-month linear path. That's fine for a learner and useless in a meeting. The map's real value is triage: a customer says "we want an AI assistant for our support docs," and in ten seconds you can place it — L1 (it's a RAG app), with L2 questions coming fast (what will it cost to serve?) and an L4 flag waving (is their data going to a third-party model?). You've turned a vague ask into three concrete workstreams before anyone's opened a laptop.

🧩 Worked Scenario: Placing a Real Ask ​

A prospect says: "We want to put a chatbot on our internal wiki so employees stop pinging the ops team." The map doing its job:

  • L1 · the ask itself — this is RAG over the wiki, the visible product; retrieval quality decides whether it's trusted.
  • L2 · the cost question — which model serves it, hosted or self-run, and per-query cost and latency at their volume.
  • L3 · usually skipped — no model training here; L3 only enters if they later fine-tune on their own data.
  • L4 · the quiet blocker — does the wiki contain HR or customer PII, and where does it go? This can stop the deal — surface it early.
Say it like this

"What you're describing is a retrieval system — the AI reads your wiki before it answers, so it's grounded in your docs, not making things up. The build is straightforward. The two questions I'd nail down early are what it costs to run at your volume, and whether anything in that wiki can't leave your environment — because that changes the architecture, not just the price."

🚨 Failure Path ​

The classic SE mistake is answering at the wrong layer. The customer asks an L4 governance question ("is our data training their model?") and the engineer answers with L1 implementation detail ("we use RAG with hybrid search"). Both true; only one is responsive. The map keeps you honest about which question you're actually being asked — so you can say "good question, that's a different layer, let me come back to it" instead of bluffing.

⚠️ Gotchas ​

  • Treating the layers as rigid — real systems blur them; the map is a thinking tool, not a standard.
  • Jumping to L1 build talk before the L4 data question — the governance answer can change the whole architecture.
  • Assuming every AI project touches L3 — most app work never trains a model.