Artificial Intelligence · Architecture

What you install, and what you maintain

Most enterprise AI proposals are quoted as one thing and deployed as nine. The licence is the small number. The bill of materials is the large one, and it arrives later.

Order or Inquire

The parts an assembled stack usually needs

A conversational assistant that can search your documents and act in JD Edwards® is not a single component. Built from parts, it typically requires most of the following — each with its own release cadence, its own security advisories and its own upgrade path.

Component	What it is there for	In BrainStorm
Chat front end	The interface users actually see	Included
Web server	Serving that interface	Included — built in
Document ingestion pipeline	Turning manuals and policies into searchable material	Included
Embedding service	Converting that material for retrieval	Included
Vector store	Holding it	Included
Retrieval layer	Finding the right passage and citing it	Included
Model gateway	Routing requests to whichever model you use	Included
Tool / MCP hosting	Letting the assistant call anything at all	Included
Identity broker	Making the answer belong to a real user	Included — the user’s own JDE sign-on
JD Edwards integration	Reaching the ERP without a shadow path	Included — the same AIS surface every AIS client uses
The language model	Producing the answer	Not included — yours

That last row is the honest one. BrainStorm does not ship a model and does not want to own yours. Point it at a local model through Ollama, or at a hosted frontier model. The choice is configuration, and it is the only external moving part.

Why the count matters more than it looks

Every component is a standing obligation

Ten components is ten changelogs to read, ten advisory feeds to watch, ten compatibility matrices to reconcile at upgrade time, and ten places for a version to drift. None of that appears in the proposal, and all of it lands on the team that already keeps JD Edwards running.

Integration seams are where things break

A failure inside one product is a support call. A failure between two products is an argument about whose it is. The more seams, the more of the second kind.

Identity is the seam that matters most

In an assembled stack the chat tool, the retrieval layer and the ERP connector each have their own idea of who the user is, and the usual reconciliation is a service account. That is the point at which the audit trail stops naming a person.

What one installer changes

One thing to install, on one server, on your own network.
One version number to record in a change request.
One vendor to call when something does not work.
One security review, covering the whole path from question to JD Edwards record.
Nothing leaves your network at all if the model is a local one.

See it working

Carrying one identity

BrainStorm carries one identity from sign-on to the JD Edwards record, because there is nowhere for a second one to appear. The evidence page shows an approval made from chat and the record JD Edwards kept afterwards, with the signed-in user’s own ID against it.

BrainStorm

Conversation, search and execution for JD Edwards, in one installer.

Local LLM AI for JD Edwards

Running the model on your own hardware where privacy, cost or data residency matters.

AI MCP Servers

The same tooling approach, available separately to other AI agents.

Composer

The governed capability layer — what the AI may do, decided in advance.

All product names, logos and trademarks are the property of their respective owners. References indicate compatibility, not endorsement or affiliation.