JM Field AI OS
Prepared for Ryan · Confidential
Infrastructure Proposal

An AI operating layer on infrastructure you already own.

The blueprint we received describes a private AI platform for J.M. Field. It is a good direction. But it was written as if we are starting from nothing, and we are not. This plan grounds the vision against the stack we already run, then starts with the work that pays back fastest and carries the least risk.

0
Standalone R740, already owned and sitting idle, becomes the AI node
0
CPU cores on that box, running its own Proxmox
GB
RAM already installed, with all-SSD storage
~$4k
Used-market parts to add GPUs and finish it

A direction, not a shopping list.

Most of the blueprint's "data" and "automation" layers are already running here. The right move is to integrate, not rebuild, and to spend effort where there is a real gap.

What the blueprint proposes

As written
  • A new GPU server as the "AI compute node"
  • Open WebUI, LiteLLM, Postgres, Redis, Qdrant, n8n, Grafana
  • RAG knowledge base over all company data
  • An "AI employee" for every department
  • A public, internet-facing portal at portal.jmfield.com

What J.M. Field already runs

Today
  • A standalone R740 sitting idle, ready to repurpose
  • An automation server that already pulls vendor invoices nightly
  • An AnythingLLM instance already serving the marketing site
  • Sage 50, Active Directory, Exchange, the Synology NAS
  • Home Assistant sensors in the warehouse

We are building on a running plant.

This is the stack the AI layer plugs into. None of it gets thrown away.

The AI node

One standalone R740

A separate server running its own Proxmox: 16 cores, 256 GB RAM, all-SSD, sitting idle today. This one box runs the AI layer. The production servers are not part of it and are not touched.

Accounting

Sage 50 US

Live on a Windows server with about 14 daily users, with data already syncing to the NAS. Stays untouched.

Automation

Invoice robots

A server that already downloads UPS, FedEx, 4over, Comcast and more every night and files them into Accounts Payable.

Identity

AD + Exchange

Active Directory and Exchange 2019, synced to Microsoft 365. The AI layer authenticates against this.

Knowledge

AnythingLLM

An AI assistant already running for the website. We extend this rather than stand up a second system.

Warehouse

Home Assistant

Temperature sensors and cameras already wired in, ready to grow into operations alerting.

Where AI actually earns its keep.

One card per department you asked about. Open each to see what fills the gap, the first step, and what to deliberately skip.

The box exists. Two parts need buying.

We pulled the live hardware report from the R740 itself. The compute, memory and storage are already strong. The only real blockers are the power supplies and the GPU mounting kit. And the GPU the blueprint recommended will not physically fit this chassis.

CPU2× Xeon Gold 6134, 16 cores
Memory256 GB
StorageAll-SSD ZFS, passthrough
Remote mgmtiDRAC9 Enterprise
NetworkDual 10 GbE
Power supplies2× 750 W, too small
GPU mounting kitNot installed
GPU fittedNone yet

GPU procurement

The blueprint said "2× RTX 3090." Those are triple-slot, 350 W open-fan cards that will not fit or cool in this 2U server. Server-grade blower cards are the right call. Pick a path:
Approx. used total$0

Sequenced by payback and safety.

Nothing internet-facing until the old, unsupported servers are dealt with. The early phases need little or no GPU and carry no risk to the accounting system.

Phase 0 · Now

Inventory and baseline

Confirm the R740 has room, capture the Sage version, and decide how financial data comes out. Done already for the hardware. Read-only, no changes.

No cost · No risk
Phase 1 · Highest ROI

Reporting and AP, internal only

Read-only Sage financial dashboards, a unified "ready to enter" view across the invoice robots, and internal document search through the existing AnythingLLM. All on the hardware we own.

Little/no GPU · Zero Sage write risk
Phase 2

Business integration

Live financial dashboards, an executive KPI view, and document search widened to marketing and operations. GPUs land here and start doing real work.

GPUs online
Phase 3

Automation and print preflight

Artwork preflight that explains errors in plain English, plus multi-step automations: invoice arrives, gets checked against the order, flagged for approval.

Department workflows
Phase 4 · Last

Cautious external access

Only after the old servers are retired and a proper isolated network is built: a single, locked-down outside service such as order status. Never a full company-data portal on the open internet.

After security hardening

Where to invest, what to skip.

The fastest way to waste this budget is to build the impressive-sounding parts first. Here is the split.

Invest

Real payback

Read-only Sage dashboards
Cash flow, AP aging and sales visibility with zero risk to the live ledger. The number-one ask from any owner.
Invoice capture, unified
Extend what already runs into one approval-ready view, with AI reading the harder PDF invoices.
Artwork preflight assistant
Catch bad print files before they cost a reprint, with plain-English explanations for customers.
Retire the old servers
The single highest-value risk reduction in the whole plan, and a prerequisite for anything external.

We pressure-tested this with three AI models.

Rather than trust one opinion, we checked the plan against three independent AI systems from different vendors. They agreed on every major call: read-only for Sage, deterministic tools for print preflight, retire the old servers first, and the 3090 is wrong for this chassis.

DeepSeek V4 Pro
1.6T parameter reasoning model

Mapped each department gap to what already exists and flagged the chatbot-per-department idea as overbuilt.

Kimi K2.7
Independent cross-check

Confirmed deterministic prepress tools beat a vision model for technical artwork checks, and that the public portal is the top risk.

Gemini Pro
Google frontier model

Surfaced the one thing the blueprint forgets: a governance plan for who owns the data and how wrong answers are handled.

The blueprint's blind spot

It is a list of technology, not an operating plan. Before any of this goes live we name who curates the data, what happens when the AI is wrong, who maintains the models, and how the knowledge base gets backed up. Skip that and you have built a fast, unmanaged liability.

Start small, start safe, start this quarter.

Three moves get us a real return without betting the business or spending much capital.