# IBM i (AS/400) Custom RPG Applications Private AI Layer Blueprint Generator

> Source: https://netray.co/tools/ibm-i-rpg-private-ai-layer-blueprint-generator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Fill in your IBM i RPG environment and export an architecture plan for a private, on-prem AI assistant: components, security controls, rollout and open questions.

## Category

on-prem-ai

## Erp

IBM i / AS400 custom RPG

## Audience

- CIO
- IT architect
- Security officer

## Tool

**Kind:** template

**Output:** document

**Fields:**

#### Company name

**Key:** company

**Type:** text

**Default:** Acme Precision Ltd

#### System version

**Key:** version

**Type:** select

**Options:**

- IBM i 7.3
- IBM i 7.4
- IBM i 7.5

**Default:** IBM i 7.4

#### Database engine

**Key:** db

**Type:** select

**Options:**

- Db2 for i (native)

**Default:** Db2 for i (native)

#### Hosting

**Key:** hosting

**Type:** select

**Options:**

- On-prem server
- Private cloud
- Air-gapped

**Default:** On-prem server

#### Named users

**Key:** users

**Type:** number

**Default:** 250

#### Sites or plants

**Key:** sites

**Type:** number

**Default:** 3

#### Main use cases

**Key:** usecases

**Type:** textarea

**Default:** Plain-language questions over order, inventory and customer files; explain what a batch job or RPG program does; summarize open orders and exceptions without green-screen navigation.

#### Data that must never leave the site

**Key:** never

**Type:** textarea

**Default:** Customer financial data, pricing files, payroll libraries and any library outside the AI views. Column masks apply to sensitive fields.

#### GPU available

**Key:** gpu

**Type:** select

**Options:**

- No GPU yet (CPU only for pilot)
- 1 x 24 GB GPU
- 1 x 48 GB GPU
- 2 to 4 x 80 GB data-center GPUs
- 8 or more data-center GPUs

**Default:** 1 x 24 GB GPU

#### Target go-live

**Key:** golive

**Type:** text

**Default:** Pilot in 8 weeks, production in 6 months

**Blocks:**

#### Context

**Body:** Company: {{company}}
System: IBM i (AS/400) Custom RPG Applications, version {{version}}, database {{db}}
Hosting: {{hosting}}
Scale: {{users}} named users across {{sites}} sites
Target: {{golive}}

Goal: add a private AI assistant over the existing system without moving data to a public cloud service and without replacing the system.

#### Use cases in scope

**Body:** {{usecases}}

#### Data that must never leave the site

**Body:** {{never}}

Control: this data is excluded from the index or processed only on {{hosting}} infrastructure, with access logged.

#### Components

**Table:**

**Columns:**

- Component
- Design

**Rows:**

**0:** Data access path

**1:** Read-only profile over JDBC (JTOpen) or ODBC to reviewed SQL views; stored procedures or IWS for logic in RPG; replica via journaling

**0:** Retrieval index

**1:** Field dictionary, spool files, manuals and program documentation indexed locally

**0:** Model serving

**1:** Open-weight model on vLLM or Ollama on {{gpu}}, on a separate x86 server

**0:** Tool server

**1:** MCP or HTTP server with named view and procedure tools; EBCDIC and packed values converted

**0:** Identity and permissions

**1:** Per-role low-privilege profiles plus row and column access control on sensitive tables

**0:** Audit

**1:** Append-only log plus Db2 for i audit journal entries for the AI profiles

#### Security controls

**Body:** No direct file writes; write-back only through RPG programs via procedures with approval; AI profiles limited to AI-specific views; RCAC masks; analytical queries on a replica during batch windows.

Hosting mode {{hosting}}: when air-gapped, models, containers and packages enter through a hash-verified transfer and no component calls an external network.

#### Rollout

**Table:**

**Columns:**

- Phase
- Scope
- Exit criteria

**Rows:**

**0:** 1. Read-only questions

**1:** Cited answers over a pilot scope for a small user group on {{gpu}}

**2:** Answers verified by process owners; no permission leaks in tests

**0:** 2. Drafts

**1:** Summaries, comparisons and drafted records a person reviews and enters

**2:** Reviewers accept drafts without major edits in agreed share of cases

**0:** 3. Approved actions

**1:** Narrow actions through the supported API with a named human approver

**2:** Every action logged; rollback path tested; sign-off from system owner

#### Open questions for IT

**Body:** 1. Which libraries and files back the screens users ask about, and is there a data dictionary?
2. Which fields hold dates or codes as numeric or CHAR values?
3. What is the batch window, and can queries run on a replica?
4. Which logic exists only in RPG and must be called, not re-implemented?
5. Who still knows the application, and how long are they available?

#### Approvals

**Signatures:**

- IT architect
- Security officer
- System owner

**Notes:** This is a planning document, not a security certification. Component choices (vector store, model size) depend on your pilot measurements; confirm API names and limits against the documentation for your exact version.

## Content

**Intro:** This generator turns a handful of facts about your IBM i (AS/400) Custom RPG Applications environment into a structured architecture plan for a private AI assistant. It covers the components, the security controls tied to your hosting choice, a three-phase rollout and the questions IT has to answer before build.

**Sections:**

#### How to use it

**Body:** Enter your system version, database engine and hosting model, then describe the use cases and the data that must never leave the site. The blueprint fills the context, components, controls and rollout sections, and exports to PDF, Word or Excel. Walk through it with your IT architect and security officer, correct the defaults, and attach it to the project request. Change the GPU field to see how the rollout row for phase one reads on different hardware.

**Bullets:**

- Be specific about restricted data: it becomes a design constraint, not a footnote
- Pick hosting first; it sets the security controls
- Keep use cases to three or four for the pilot
- List the people who can approve write-back

#### What the blueprint assumes

**Body:** The design is the same on every on-prem system: a read-only copy or supported API for data access, a retrieval index that carries permissions, a tool server that exposes only approved read actions, an open-weight model on vLLM or Ollama, and an audit log. The real cost is the semantic layer over short field names, not GPU hardware. The components table names the access path for this system, but you must verify the object and API names for your exact version.

**Bullets:**

- Read-only by default
- Permissions mirrored from the system
- Write-back only through supported APIs with human approval
- Everything logged

#### Reading the rollout table

**Body:** Phase one proves the data and permissions work: cited answers over a small scope. Phase two adds drafts that a person reviews. Phase three allows narrow, approved actions through the system's own API. Do not skip phases: most failures come from skipping permission testing in phase one. Exit criteria are examples; replace them with numbers your process owners agree to, such as the share of answers verified correct in a sample of fifty questions.

#### Common mistakes

**Body:** Teams often size the model before they have defined the use case, grant the assistant a powerful service account, or index everything to avoid deciding scope. Another is leaving the open questions unanswered until build starts. Small user base, high question rate. Use the open-questions section as the agenda for your first IT workshop and assign an owner to each item.

**Bullets:**

- Powerful service account for convenience
- No permission tests with restricted users
- Indexing without scope decisions
- No owner for open questions

**Cta:** Netray designs and runs private AI layers on IBM i (AS/400) Custom RPG Applications with your IT team. If you want a second opinion on your blueprint, we can review it with you.

**Faq:**

#### Does this document commit us to buying anything?

**Answer:** No. The generator runs in your browser and exports a plan you can share internally. It is a starting point for architecture and security review, and every default can be changed.

#### Is the plan valid for air-gapped sites?

**Answer:** Yes. Choose the air-gapped hosting option and the controls section adds transfer and network isolation rules. All components in the table run inside the closed network with no external calls.

#### Why does the plan avoid giving model sizes?

**Answer:** Model size and GPU count depend on concurrency, context length and answer quality targets. Measure them in the pilot, then use the sizing calculator for ranges rather than committing to a number early.

#### Can we reuse the plan for another system?

**Answer:** The structure transfers, but the data access path, permission model and open questions are specific to each system. Generate a separate blueprint for each platform and compare them.

## Search terms

ibm i rpg ai architecture, ibm i rpg private llm plan, on-prem ai blueprint template, air-gapped ai architecture document, private ai layer architecture template, ai assistant security controls template

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json