The Enterprise AI Stack

An AI feature in a serious organization is never just "call a model." It sits on a stack of layers, each owned and reasoned about by the architect. Understanding these layers is what separates a demo from a system you can run, secure, and defend in a review.

The layers

From the bottom up, a typical enterprise AI feature looks like this:

1. Data & sources — the documents, records, and content the AI is allowed to see. In a .NET shop this is usually SQL Server, plus content systems like SharePoint or a CMS.
2. Ingestion & indexing — a service that pulls approved content in, splits it into chunks, and turns each chunk into an embedding for search. This is a pipeline, not a one-off.
3. Retrieval — given a question, find the most relevant chunks (vector + keyword search).
4. Inference — the model call itself: embeddings (often local) and generation (local or cloud), behind an interface so the provider can change.
5. Orchestration — the application logic that wires retrieval, prompting, validation, and guardrails together into a use-case.
6. Presentation — the API and UI the user actually touches.
7. Cross-cutting — security, observability, audit, and governance, present at every layer.

Why the stack matters

Each layer is a place to enforce a quality goal. Security lives at the data and inference layers (what can the AI see; where does data go). Correctness lives at retrieval and orchestration (is the answer grounded; was it validated). Cost lives at inference (local vs cloud). Observability spans all of them.

When something goes wrong — a hallucinated answer, a leaked record, a runaway bill — the architect's first move is to ask which layer failed. A platform without clear layers gives you nowhere to look.

How LyraLearn maps to the stack

This very platform is built on these layers, which makes it a working reference:

• Data: SQL Server (lessons, knowledge-base documents, embeddings).
• Ingestion: a background importer that chunks and embeds lesson content.
• Retrieval: SQL Server 2025 native vector search.
• Inference: local embeddings + local/cloud generation behind interfaces.
• Orchestration: the AI Tutor service (retrieve → grade → ground → answer).
• Presentation: this web app and its API.
• Cross-cutting: audit logs and an evaluation record for every AI call.

As you move through the course, you'll build each of these layers yourself.