Reference design

On-premise AI architecture, one site or many

How InitD designs a private AI installation: the hardware and network, how your devices and applications connect, where the data stays, and how models improve on your own data. On your premises, in your data centre, or both. Every installation is tailored to its site.

Discuss your setup Back to Private AI →

Benefits

Your data never leaves your control

Confidential records, recordings and client files stay on your premises or in your data centre, supporting your data-protection and client obligations.

Keeps running when the internet doesn't

Your AI carries on through outages and catches up when the connection returns.

Costs you can plan

Hardware you buy once, or usage billed to your own cloud account. No per-request fees to a public AI provider.

Yours to keep, no lock‑in

You own the hardware and the accounts, and your team gets everything needed to run it without us.

For your IT team

Raw data processed locally

Video, audio and documents are stored and processed on your infrastructure.

Outbound-only management

Encrypted outbound connections only. No inbound firewall rules at any site.

OpenAI-compatible private API

Applications built for public AI services can often switch with a configuration change.

Connects to your inputs

Cameras, microphones, scanners, documents and business systems.

On-site fine-tuning

Models adapted on your own data, tested against the current model before going live.

Built-in resilience

RAID storage, verified transfers and staged rollouts with automatic rollback.

What You Can Build

One private AI stack, many uses

Four illustrative examples. Each starts with what changes for the business, then shows how it works.

Monitoring

Spot the moments that need a person

Cameras and microphones flag incidents as they happen, so nobody has to watch screens all day, and the footage stays on infrastructure you control.

For example: safety on a factory floor, quality checks on a line, security after hours, care settings.

  • Incidents flagged in seconds, with the clip attached
  • A short written summary of what happened
  • Raw video and audio kept on your infrastructure, deleted on your schedule

How it works

Your inputsCameras, microphones, sensors
CamerasMicrophonesSensors
Edge devicesDetect events on the device; record only when something happens
Private AIVision-language model reviews each clip and writes a summary
GPU computeLocal storage
Where it landsAn alert with clip and summary in Teams or your app

Typical starting setup: edge devices per camera, with two linked GPU units on site or GPU servers in your data centre, and storage sized to your retention period.

Documents

Search and draft from confidential documents

Staff find answers across contracts, case files and policies in seconds, and get a first draft to edit, without a single document going to a public AI service.

For example: law firms, finance and accounting teams, HR, healthcare administration.

  • Answers with links to the source document
  • First drafts in your own templates and tone
  • Access follows the permissions you already have

How it works

Your inputsFile shares, document systems, scanned paper
File sharesDocument systemsScans (OCR)
Local search indexDocuments indexed on your infrastructure, respecting who can see what
Private AILanguage model answers with citations; fine-tuned on your templates
GPU computeFine-tuning
Where it landsA secure chat in your intranet, Teams or document system

Typical starting setup: one GPU unit on site, a GPU server in your data centre, or capacity in your own cloud account.

Operations

An assistant that knows your operations

Technicians and frontline staff ask questions on a phone, tablet or in Teams and get step-by-step answers from your manuals, procedures and maintenance history.

For example: manufacturing, facilities management, clinics, field service.

  • Fewer calls to the one person who knows
  • Photos of equipment understood, not just text
  • Keeps working on site when the internet does not

How it works

Your inputsManuals, SOPs, maintenance logs, equipment photos
ManualsMaintenance logsPhotos
Local knowledge baseYour documents indexed and kept up to date
Private AILanguage and vision models answer from your own material
GPU computeOffline-capable
Where it landsTeams, a tablet app or your maintenance system

Typical starting setup: one GPU unit on site, so answers keep coming during internet outages.

Your applications

Add AI to the systems you already run

Your in-house applications, ERP or CRM gain summaries, classification and drafting, without sending customer data to a public AI service.

For example: any business with in-house software or a development team.

  • AI features inside the tools staff already use
  • Your developers point existing code at your own models
  • Usage, quotas and an audit log you control

How it works

Your inputsYour application and its data
In-house appsERPCRM
Private AI APIOpenAI-compatible endpoint, with authentication and quotas
Private AIModels selected and fine-tuned for your tasks
GPU computeFine-tuning
Where it landsNew AI features inside your own application

Typical starting setup: GPU servers in your data centre or capacity in your own cloud account, scaled to how many requests your applications make.

Infrastructure

How is an on-premise AI site laid out?

Whatever the use case, each installation gets the same building blocks: GPU compute sized to the workload, local storage and a dedicated switch, connected to the network you already have, on your premises or in your data centre. Inputs, from cameras to business systems, feed in over your own network. One outbound-only connection links each site to central management.

More sitesThe same stack at each location, or devices only
Your data centreServes many sites from one place, with enterprise power and cooling
+ more sites as you grow

Built to be qualified, not assumed

Network speeds, cables and ports are proven on a test unit before anything is bulk-ordered.

  • Compute: compact GPU units for offices, rack-mount GPU servers for data centres; one unit for many workloads, linked units for larger models, a small cluster for heavy, concurrent use.
  • Storage: sized to your retention period, on its own fast link.
  • Edge devices (where used): managed appliances with no user logins, updated only in maintenance windows.
  • Your network: uplink capacity, private links between sites and Wi-Fi coverage checked in the survey.

Resilience

What happens if the internet goes down, or a drive fails?

The system keeps working. Detection, recording and AI processing run locally, and events sync to central management when the connection returns. Storage uses RAID-6, so it survives two failed drives.

Built in

  • Offline operation: local processing continues; pending events sync on reconnection.
  • RAID-6 storage: tolerates two drive failures. RAID protects against failed drives; it is not a backup.
  • Verified transfers: every file checksummed, with automatic retry and resume.
  • Staged rollouts: updates go one device, then one site, then everywhere, with automatic rollback.
  • Health dashboards: device heartbeat, compute temperature, storage capacity and network links.

Added when uptime demands it

  • Failover compute: an extra unit that takes over automatically.
  • Replicated storage: a second unit or off-site backup for your recordings.
  • UPS and dual power: rides through power cuts and shuts down cleanly. Standard in most data centres.
  • Spare units: for the fastest hardware swaps.

We size redundancy to what an outage would cost you, and say plainly what each option does and does not cover.

Data Sovereignty

Where does the data live, and who owns it?

On infrastructure you control. Raw video, audio and documents are stored and processed on your premises or in your data centre; central management only sees events and device health, over encrypted outbound-only connections.

Accounts in your nameAny cloud accounts are yours from day one, so nothing needs handing over later.
Your retention rulesYour data protection officer sets how long recordings are kept and when they are deleted.
Held when it mattersRecordings tied to an open issue can be kept beyond normal retention.
Your sign-inSingle sign-on with your directory, roles and MFA for administrators.
Every action loggedAudit trails for access, recordings and changes.
Hardware you ownOn-premise and data centre equipment is your asset; handover includes full admin access.

Device Integration

What devices and data sources can connect?

Almost anything that produces images, audio or data. Cameras, microphones and sensors are common starting points, alongside scanners, machines and the CCTV you already have. Where it helps, small edge computers sit beside them and process data locally, so only what matters is recorded or sent on.

Works with what you have

Existing IP cameras and CCTV, microphones, scanners, sensors and machines, connected through standard interfaces: for example video streams (RTSP, ONVIF), industrial protocols (MQTT, OPC UA) and APIs.

Processing at the edge

Small computers beside your devices detect events locally, with thresholds you configure, so bandwidth and storage go to what matters.

Captured when it matters

Recording or data capture triggered by an event, a schedule or a person, from your own app or Microsoft Teams. Every action is logged.

Managed devices

Each device is enrolled with its own revocable certificate and updated centrally.

Using something not listed here? Tell us what you have. Compatibility is checked in the site survey and proven on the test unit before anything is ordered.

Application Integration

How do our applications use the AI?

Through a private API that follows the OpenAI format, so applications built for public AI services can often switch with a configuration change. AI results flow into the systems and tools your staff already use.

Connect what you already run

  • Your systems: management systems, internal portals and databases.
  • Microsoft Teams: alerts, commands and an assistant where staff already work.
  • Access control: roles, quotas and an audit log on every request.

AI added to existing applications

  • Vision on your infrastructure: a vision-language model reviews flagged footage and writes a short summary.
  • Context-rich alerts: the right person is told what happened, with the clip attached.
  • An assistant on your data: staff ask questions and get answers from your own records.
  • Follow-up suggestions: practical next steps after an event, reviewed by a person.
# Your application, pointed at your own models
client = OpenAI(base_url="https://ai.yoursite.internal/v1")

client.chat.completions.create(
    model="your-private-model",
    messages=[{"role": "user",
               "content": "Summarise today's flagged events"}],
)

LLM Fine-Tuning

Can we fine-tune the models on our own data?

Yes. Models are adapted on your own infrastructure, on your own data, in scheduled windows outside working hours, and only go live after they beat the current model on agreed tests.

Light, not from scratch

Small adapter layers (LoRA) trained on top of an open model, so training fits on your own hardware.

Tested before it ships

Each new version is evaluated against the current one and regression-tested, with rollback if anything slips.

Sites learn from each other

With several sites, each trains on its own data and shares only the model update, encrypted and signed. Recordings never leave.

Fine-tuning is also available for cloud deployments: see Enterprise AI.

Delivery

From site survey to handover

  1. Step 1: Site survey

    Space, power, network and what you need AI to do, on site or in your data centre.

    A clear plan

  2. Step 2: Test unit

    One unit proven in your environment before anything is bulk-ordered.

    No surprises at scale

  3. Step 3: Install

    Hardware installed and connected to your systems and network.

    A running system

  4. Step 4: Benchmark

    Accuracy, speed and capacity measured on your own data.

    Measured, not promised

  5. Step 5: Handover

    Agreed tests passed, team trained, documentation delivered.

    Your team runs it, or we do

Want this designed for your sites?

Tell us how many locations you have, what you need AI to see and hear, and where the data must stay. We will design the installation with you and give you a fixed figure.

Discuss your setup