For your technical reviewer

How it’s built.

Before a school or small business trusts an AI system with its records, someone technical usually gets a phone call — a board member who ran networks, a consultant, a friend who “does computers.” Good. This page is for whoever you ask to check our work.

Design goals

Four decisions, made before any hardware is ordered.

One machine

Inference, chat interface, document index, and workflow automation all live on a single box in your office or network closet. Fewer parts, fewer failures, one thing to point at.

Private by default

Every service binds to your local network only. No cloud APIs, no per-token costs, no data-processing agreements to negotiate — the data has nowhere to go.

Economical

Enterprise-grade hardware and free, open-source software. A one-time cost instead of a subscription that grows with use.

Maintainable

Patches, model upgrades, and workflow fixes without a site visit — and without opening a single inbound port. That tradeoff gets its own section below.


The machine

A workstation-class machine, sized for the job.

Platform

Workstation-class tower · Xeon-based · 12 or more cores.

Memory

64–128 GB ECC.

GPU

RTX 3060-class, 12 GB of video memory — sized for 7–14B-parameter models with retrieval context to spare.

Storage

NVMe for the operating system · SSD for models · a high-capacity drive for documents, the search index, and backups — plus a second backup target.

Power

UPS-backed · roughly the draw of a large household appliance.

Noise

Office-closet quiet — a workstation, not a rack server.

Every machine is burned in before deployment — memory tests, GPU stress, drive health checks — because the box has exactly one job: to be boring.


The stack

Containers, pinned versions, deliberate upgrades.

Everything runs as containers from a single, version-pinned configuration on a long-term-support Linux server release. Upgrades happen deliberately, in maintenance windows — never automatically, never latest. Stability beats freshness in a school office. Automatic security patching stays on for the operating system, with the GPU driver pinned so an update never quietly breaks inference.

Inference server

A local inference engine (Ollama-class) runs the language and embedding models on the GPU. Nothing calls out.

Chat + doc index

The interface staff actually use: a web page on the internal network with a private document knowledge base built in, so answers are grounded in your own files.

Workflow engine

Where the agents live — letter engines, form-chase campaigns, deadline ticklers, and validation runs, each built as an auditable workflow.

Speech-to-text

A Whisper-class transcription service for voicemail and meeting audio. Entirely local; runs on the CPU when the GPU is busy.

Parsing / OCR

The scanning pipeline that turns paper and PDFs into searchable, classified records — the feedstock for everything else.

Monitoring

A service-health dashboard reviewed on every maintenance pass, so problems are found by us, not by your front office.

Reverse proxy

Optional: friendly internal names — assistant.school.lan — with LAN-internal TLS.


The models

Scoped, quantized, and role-specific.

Main assistant

A 7–14-billion-parameter instruction-tuned model, quantized to sit comfortably in GPU memory with room left for document context.

Fast lane

A smaller model handles classification and routing tasks where speed matters more than depth.

Embeddings

A dedicated embedding model powers the document index.

Each role gets a named model preset carrying its own system prompt, parameters, and a context window capped to what the task actually needs — the assistant that answers policy questions is deliberately not the same configuration that drafts letters. Scoping this way keeps behavior predictable and GPU memory honest. Model choices are re-evaluated at every deployment; the field moves fast, but the architecture doesn’t have to.

The tradeoff we volunteer

A true air gap and remote management are mutually exclusive.

Anyone who tells you otherwise is selling something. So we offer two postures, explain both honestly, and let you choose.

Posture 1

Fully air-gapped

No outbound connectivity, ever. Maintenance happens on-site — or on a scheduled maintenance day where the machine briefly pulls updates and is re-isolated.

  • Highest assurance available
  • Higher maintenance labor, by design
  • Some records deserve exactly this

Either way, you hold the plug. The handoff document literally includes how to pull it.

Hardening & backups

Boring on purpose.

Hardening baseline, either posture

  • SSH by key only — no passwords, no root login, automatic lockout on repeated failures.
  • Default-deny firewall; staff-facing services allowed from your LAN subnet only.
  • Containers on an isolated internal network — only the interfaces staff use are published at all.
  • Operating-system security patches applied automatically; the GPU driver held pinned.

Backups that get restored

  • Nightly backup to a second drive: database, documents, workflows, and configuration — models excluded, since they’re re-downloadable.
  • Weekly encrypted copy to a rotated USB drive kept by your office manager — an off-site equivalent that never touches the internet.
  • A full restore is tested at deployment, and again every quarter after.

The cadence

What maintenance actually looks like.

Automatic

Security patches, nightly backups, uptime monitoring.

Monthly · remote

About an hour: dashboard review, disk space, backup logs; pinned container updates applied if warranted.

Quarterly · remote

About two hours: restore test, model performance review, workflow tune-ups, document re-index if the files have churned.

Annually · on-site

Optional: dust the machine, verify the UPS battery, refresh the documentation, sit down with your staff for feedback.


Cost posture

One-time, then quiet.

The hardware is a one-time purchase — enterprise-grade equipment, sized to the job and burned in before it ships. The software stack costs nothing, now or later, and the power draw runs about the cost of a large appliance. There are no per-seat licenses, no per-token fees, and no bill that grows because your staff actually use the thing. Real numbers come in the proposal, sized to your building.


What we hand over

Yours, not a mystery box.

  • Acceptance testing you can watch: retrieval quality validated on small document batches before bulk loading, and the assistant’s refusal behavior tested on out-of-scope questions — it should decline to speculate, and we prove it does.
  • Staff orientation: thirty minutes, on-site, with the people who’ll use it daily.
  • A plain-language handoff document: what’s running, where your data lives, how to reach us — and how to pull the plug.

Have your skeptic read this page. Then have them call us.

We like talking to the person you trust to check our work — that conversation is where good deployments start.

Start a conversation →