INTERpriseAI
Private agentic architecture  ·  Deployed inside your network  ·  Sydney, Australia

We install AI inside your business.

Private agents that run on your own hardware, answer from your own data, and cost nothing per question, per seat, or per user.

Layer 01
Identity
Who is allowed in, and how
Layer 02
Gateway
Which model, at what cost
Layer 03
Agents
Work that actually completes
Layer 04
Evidence
Traces, logs, uptime, audit

What we build

Three practices
Practice 01

Agentic architectures

Retrieval and reasoning systems wired into the databases, ERPs and schedulers a business already runs on — with failure modes designed in, not discovered later.

  • RAG over live operational data
  • Failure-aware response design
  • Human-in-the-loop approval paths
Practice 02

Automation & orchestration

Workflow automation that carries production discipline: versioned, monitored, alerting on its own failures, and reviewable line by line before it ever touches your data.

  • Workflow and API orchestration
  • Scheduled and event-driven pipelines
  • Operator alerting and escalation
Practice 03

Sovereign & private AI

For regulated and defence-supply-chain work: models that run on your own hardware, on your own network, where no prompt leaves the building.

  • On-premises models, no GPU required
  • Active Directory and LDAP integration
  • Essential Eight ML2 alignment
  • Data-residency by architecture

The reference architecture

Full stack  ·  Observability at every layer
Edge
Reverse proxy & TLSOne front door, managed certificates, split DNS
single ingress
Identity
Directory federation & brokered SSOActive Directory and LDAP integration, partner tenants, conditional MFA
least privilege
Gateway
Model routing & budget controlScoped keys, per-application aliases, spend ceilings
cost ceiling
Reasoning
Agents, retrieval & workflowVector store, orchestration, conversational surfaces
no data egress
Evidence
Tracing, metrics & security telemetryEvery prompt traced, every service watched, every event retained
full audit trail

We deliver the whole stack, not a layer of it — and every layer reports on itself, so the platform can be proven to be working rather than assumed to be. We do not publish the bill of materials. Components are selected per client against their regulatory and network constraints, and disclosed under engagement; naming the internals of a client's platform on a public page is reconnaissance we are not willing to hand out.

The hardware question

Local models

No GPU.

For everyday operational tasks, a locally hosted model does the work most people currently send to a public AI service — with the difference that nothing leaves your network.

Footprint
A server VM you already own
Ordinary CPU and memory on existing virtualisation. No accelerator purchase, no new rack, no data-centre conversation.
Escalation
Add a GPU only when the work earns it
Larger models and heavier workloads get hardware when there is a measured reason. The architecture does not change when they do.
Routing
Local by default, cloud by exception
Sensitive work stays on the local model. A task that genuinely needs a frontier model is routed deliberately, logged, and costed.

Inside the platform

Three surfaces, one stack
Surface 01

The staff portal

One address for everything AI in the business. Tiles a person is not entitled to stay visible but inert, so staff can see what exists and request access — rather than discover a wall.

Built for the business sponsor
https://ai.kestrelindustries.com.au
j.harding

AI PLATFORM

Platform is healthy
Assistants
Ask Kestrel
General assistant
Scheduling agent
Production planning
Document search
Policies and drawings
Quality analyst
Non-conformance review
Workspace
Knowledge base
Company documentation
Manage sign-in
Profile & 2FA
Data curator
Training samples
Operations
Platform administration
Surface 02

The operations launchpad

Every console behind the platform in one gated page, with a live health indicator polling the status plane. Delivery is the easy part — this is the surface that proves the thing is being run.

Built for the IT manager
https://admin.kestrelindustries.com.au
platform-admin

ADMIN LAUNCHPAD

All services up
Identity & access
Directory federation
login.kestrelindustries.com.au
Access policy
auth.kestrelindustries.com.au
Session audit
audit.kestrelindustries.com.au
Perimeter
10.40.0.254:8080
Observability
Dashboards
metrics.kestrelindustries.com.au
Prompt tracing
trace.kestrelindustries.com.au
Uptime
status.kestrelindustries.com.au
Security events
10.40.0.22:9090
Surface 03

The agent at work

An answer grounded in the company's own operational data, with the source named and the limits of the answer stated. When it cannot know something, it says so instead of inventing it.

Built for the person doing the job
https://plan.kestrelindustries.com.au
j.harding
JH
Which orders on line 3 are at risk of missing their promised date this week?
K
Four orders on line 3 are currently forecast past their promised date.

SO-48117 is the most exposed — it is queued behind a changeover and has no float left. SO-48122 and SO-48130 lose their buffer only if the changeover runs over its standard time. SO-48145 is at risk from a material receipt that has not been booked in.

I can see the schedule and the order book, but not the inbound delivery status — the receipt date for SO-48145 needs confirming with purchasing before you treat this as final.
Source · production schedule, order book · read 14 minutes ago
Ask a follow-up question

Representative interfaces · fictional company, fictional data · no client system depicted

How we work

Four commitments
01

Architecture before tooling

TOGAF-grounded design. We establish the target state and the constraints — regulatory, network, budget — before a single product is chosen.

02

Production discipline from day one

Backups before edits, controlled restarts, a stated rollback path for every change. There is no "it's only the pilot" phase.

03

You get the documentation, not a dependency

Every engagement ends with a set of documents your own team can work from. Lock-in is a business model, not an outcome.

Architecture designAs-built recordDeployment runbookOperating proceduresRecovery & rollback planAccess & group model
04

Compliance is designed in

Defence Industry Security Program and Essential Eight Maturity Level 2 obligations shape the architecture from the first diagram, not the final audit.

Who this is for

And who it is not
A good fit

This is for you if

  • Your industry, your insurer or your customers impose obligations on where data lives and who can reach it.
  • You already run your own servers, or you are willing to, and you would rather own the thing than rent it per seat.
  • Your team is doing work by hand that a system should be doing, and you can name it.
  • Someone internal will hold the keys afterwards — you want documentation, not a retainer you can never leave.
Not a fit

This is not for you if

  • You want AI to replace your team. It will not, and we will not sell you that.
  • You want the cheapest option rather than the right one. A public subscription is cheaper on day one, and that is a legitimate choice.
  • Nobody internally owns IT. Something that runs inside your business needs someone inside your business who cares that it is running.
  • You need it live next week. Architecture done properly takes longer than that.

Still sceptical? Good.

Three honest answers
Objection 01

"We tried AI and it made things up."

So does an unconstrained model with no grounding. Ours answer from your documents and your databases, name the source, and state plainly when a question falls outside what they can see. The failure modes are designed and tested before anything reaches a user.

Objection 02

"It will become another thing we depend on you for."

Everything we build is handed over documented, down to the configuration files and the rollback path. If you replace us, the platform keeps running and your team has what they need to run it. That is a deliberate commercial position, not an accident.

Objection 03

"Our data cannot leave the building."

Then it does not. The models run on your hardware, on your network, behind your perimeter. This is the constraint we design for first, not a feature we bolt on afterwards for regulated clients.

Reference engagement
A regulated Australian manufacturer, from zero AI capability to a governed platform serving staff and partner organisations.
18
Services governed
2
Identity providers brokered
ML2
Essential Eight target
0
Prompts leaving site

How to start with private AI architecture

No obligation

Bring us a process you think an agent should own. We will tell you honestly whether it should, what it would take, and what it would cost to run.