Private & Sovereign AI · Perth, Western Australia

Private AI agents

for operations where

loose answers are not acceptable.

Airon Systems builds governed AI workflows for the businesses that win work by tender and answer for safety on real sites — serious operations, no internal AI team required. Source-backed answers. Human approvals where judgement matters. Standing rules where action is repeatable. Audit logs by default.

Map a private AI pilot→Request an inference audit→See Gauge and Marshal→

▸More tenders out the door with the same estimating team

▸Pricing backed by your delivered-job history

▸Safety response from the cameras you already own

▸Incident paperwork drafted from the event log

Not a chatbot

Operational AI agents, not chatbots.

Most enterprise AI stops at chat: a text box that answers questions and leaves the work exactly where it found it. Useful for drafting an email. Useless for a tender due Friday, a person down on a deck, or an auditor asking who approved what.

airon agents are built the other way around. They connect to the systems where operational work already lives — tenders, rate cards, project history, SWMS, incident logs, rosters, work orders — and carry a task from trigger to outcome, with a human wherever judgement is required.

Chatbots answer. Agents complete work.

An estimating agent doesn't summarise the tender — it reads the pack, compares past projects, checks current rate cards, flags margin risk, asks the questions an estimator would, and drafts the estimate with its evidence attached.

Connected, permissioned, logged.

Agents act through governed tools with explicit permissions, against your evidence — documents, records, registers — with approval gates where they matter. Every read, action, and decision lands in an audit log, so “what did the AI do?” is a report, not an investigation.

Built for the operators the giants overlook.

Enterprise AI assumes an enterprise: data teams, integration budgets, eighteen-month rollouts. airon is built for companies that price work off rate cards and answer for safety on real sites — without an internal AI team. We bring the build and the governance. Your people keep the judgement.

The Control Model

Five controls. Every agent.

Every airon workflow runs inside the same control frame. It's the difference between an impressive demo and a system an owner can stand behind — and an auditor can check.

01

Private data

Deployed so sensitive records — pricing, safety data, contracts — can stay inside your environment, through to on-premise and air-gapped options.

02

Governed tools

Agents touch systems through explicit, permissioned tools. What an agent can read and do is defined up front, not discovered later.

03

Human approvals

Where judgement matters — pricing calls, incident escalation — the agent stops and asks. Nothing goes out without sign-off.

04

Standing rules

Where action is repeatable and pre-agreed, agents act on standing rules and report back. You set the rules; they do the running.

05

Audit logs

Every detection, read, decision, and action is timestamped by default. Evidence you can hand to an auditor, not a mystery to reconstruct.

Working Examples

Watch Gauge and Marshal work.

Two agents from our private sandboxes, recorded as they ran. The data is fictional; the orchestration is real — source-backed drafts, human decision points, standing-rule actions, and an audit trail building in real time.

Construction estimating agent

Gauge

Gauge reads an 84-page tender, indexes it against project history, checks current rate cards, flags latent-condition and margin risk, asks the questions an estimator would — then drafts a priced estimate with every line traceable to a source.

  • ▸Tender to draft estimate with evidence attached
  • ▸Margin risk flagged from delivered-project history
  • ▸Human sign-off before anything leaves the building

Recorded from the live sandbox · fictional sample data

Safety & incident response agent

Marshal

Behind video analytics on existing cameras, Marshal turns a detection into a response: it reads the conditions, asks a human only for the calls that need judgement, runs standing-rule actions, and builds the incident record as it happens — then drafts corrective actions and updated watch rules.

  • ▸Detection to dispatched response in seconds
  • ▸Human approvals for judgement; standing rules for the routine
  • ▸Incident records drafted from the event log, not from memory

Recorded from the live sandbox · fictional sample data

The interactive versions run behind private access for clients and reviewers — decisions, debriefs, and the data library included. Request sandbox access →

Use Cases

Where agents earn their keep.

Not every workflow needs an agent. These are the ones where the maths already works: high document load, real deadlines, margin or safety on the line, and a clear human owner for every decision that matters.

01

Construction estimating

Tender packs read against project history and live rate cards. Comparable-project pricing, prelims benchmarks, latent-condition flags, and margin scenarios — drafted with evidence for an estimator to sign.

  • Tender intake and triage
  • Rate-card and subcontractor pricing
  • Margin-risk flags from delivered projects
  • Draft estimates with source trails

02

Safety & incident response

Detections from the cameras and systems you already run, turned into governed response — welfare checks, dispatch, notifications, and the paperwork that follows.

  • Camera and alert triage
  • Response fan-out with approval gates
  • Incident records drafted from the event log
  • Corrective actions and watch-rule updates

03

SOP & compliance retrieval

Source-backed answers from SWMS, SOPs, site procedures, training records, and compliance registers — with citations and access controls, designed to reduce hallucination risk.

  • SWMS and SOP lookup with citations
  • Induction and training support
  • Compliance register queries
  • Role-based document access

04

Operational knowledge systems

Private document intelligence over the records that actually run the business — retrievable in seconds, governed by role, and deployable so sensitive data stays inside your environment.

  • Project and maintenance history
  • Contractor and training records
  • Approvals and audit trails
  • Private, permissioned retrieval

Deep dives: Private AI for commercial builders →·Private AI for civil contractors →

The Pilot

Four to six weeks. One workflow. Real numbers.

Operational companies don't need more demos. A private AI pilot puts a governed agent on one narrow, high-value workflow — with your documents, your rules, and success metrics agreed before we build.

One narrow workflow

Tender triage, incident response, SOP retrieval — one job the agent can carry end-to-end, on your documents and your rules. Not a platform rollout.

Success metrics agreed first

Cycle time, response time, rework, coverage — measured against your current baseline and agreed before anything is built.

Your controls from day one

Human approval gates, permissions, and audit logs are in the pilot itself, not promised for later. Private deployment options where data can't go to generic cloud tools.

A clean decision at the end

Scale it, refine it, or stop. Either way you decide on evidence — measured results against the metrics you set.

Map a private AI pilot→

The Path

From audit to operations.

Each step is scoped from what the previous one proves. No open-ended programs, no platform bets before the evidence is in.

01

Inference Audit

Map the workloads, data boundaries, systems, and constraints. Decide where an agent is worth piloting — and where it isn't.

02

Pilot Workflow

Four to six weeks on one narrow workflow, with agreed success metrics, approval gates, and audit logs from day one.

03

Private Deployment

Move what worked into your environment — through to on-premise and air-gapped shapes — with permissions, logging, and integrations hardened.

04

Managed AI Operations

Keep it sharp as the business changes: monitoring, governance, model updates, and new workflows — run with your team, or for it.

Architecture

Cloud-to-edge. On your terms.

Not every workload needs to run in the same place. Airon designs enterprise inference topology across airon appliances, private model runtimes, approved retrieval stores, audit controls, and hybrid private AI paths that put your sensitive work on-premise and air-gapped, while everyday tasks move at cloud speed — so you get sovereignty where it matters and reach where it pays.

Air-Gapped

Sovereign Tier

On-premise, air-gapped AI systems for the work you can't risk sending offsite — safety data, compliance documents, contracts, and pricing. Fully isolated from the public internet, from devices on-site through to appliances and servers in your data hall — deployed so your most sensitive records stay inside your environment.

Isolated Routing

Secure Cloud Tier

Routes everyday, low-risk tasks to approved cloud AI services through isolated network paths. Sits alongside the sovereign tier, so your team gets the flexibility of public AI tools without breaking the air gap on the sensitive work.

Trust & Proof

Built for due diligence.

No procurement portal required. Airon is built for owners and directors who need more than a demo — company context, deployment assumptions, security posture, and pilot fit go on the table in the first conversation.

Company

Perth based. Australian focused.

Airon Systems is based in Western Australia and builds around the data, security, and operating realities of Australian mining, construction, and transport teams.

Security Posture

Sovereignty before automation.

Every engagement starts by classifying what must stay on-premise, what can route to approved cloud tools, and where audit trails need to exist.

Delivery Control

Scoped from audit to handover.

Recommendations are framed around use case value, deployment shape, data readiness, access control, logging, maintenance, and the people who will operate the system.

Qualification Pack

Proof before pilot.

Company details, delivery-team profiles, security notes, deployment assumptions, and pilot prerequisites can be shared during qualification.

In Plain English

The vocabulary, decoded.

We use a few terms a lot. Here's what they actually mean — so you can read the rest of this page (and any AI vendor's pitch) with full confidence.

Private AI
AI models you run inside your own systems, on your own data. Nothing is sent to a third-party provider, and nothing about your business trains someone else's model.
Sovereign AI
Private AI that's also built, hosted, and governed inside Australia, on hardware and software you own — so it stays under Australian law and out of foreign jurisdictions.
AI agent
Software that uses a model to complete a task, not just answer a question: it reads approved sources, calls permissioned tools, asks for approvals, and logs what it did. The difference between chat and work.
Standing rules
Pre-agreed actions an agent may take without asking — radio a worker, raise a work order, hold an interlock — always inside set permissions, always logged. Humans keep the judgement calls.
On-premise AI
AI that runs on servers physically inside your office, site, or data hall. The opposite of cloud AI, where the model runs on someone else's machines.
Air-gapped
A system with no connection to the public internet. Data physically cannot leave the building, which makes it the right call for safety records, contracts, and pricing.
Inference
The live answer-generation step. Every time someone asks the AI a question, that's one inference. Your “AI bill” is essentially a count of these.
Inference Topology
The enterprise layout that decides where AI runs: local appliances, model servers, retrieval stores, routing, policy controls, audit logs, and approved cloud paths.
RAG (retrieval-augmented generation)
Instead of guessing from memory, the AI looks up the relevant passage in your documents first, then writes its answer from that source. The result: fewer hallucinations, traceable answers.
Fine-tuning (LoRA / QLoRA)
Teaching an existing model your specific terminology, procedures, and edge cases — without retraining the whole thing from scratch. Cheap, fast, and reversible.

Industries

Built for operational scale.

Mining & Resources

Mining & Resources

Private AI systems for remote mining operations. Instant answers from your SOPs, safety procedures, and compliance documents — running on-site, with no internet needed — so new starters and crews get accurate guidance the moment they need it.

Construction

Construction

Private AI for tender and contract-heavy work. Read complex tenders in minutes, surface risks and scope traps, and cross-check against past bids — all on your own systems, so your pricing and IP never leave the business.

Transport & Logistics

Transport & Logistics

Private AI workflows for fleet operators and logistics networks. On-premise route analysis, compliance lookups, and freight risk reviews — without handing operational data to a third-party cloud platform.

Field Briefs

Download the working notes.

Two short PDFs give internal sponsors something concrete to circulate before an inference audit or airon pilot conversation.

Start Here

Start with an Inference Audit.

We map how the work actually runs — tenders, incidents, SOPs, approvals — where the data boundaries sit, and which single workflow a pilot should prove first. You get a deployment plan with numbers attached, and a clear read on where private AI is worth your money — and where it isn't yet.

Request an inference audit→Map a private AI pilot→

Start a Conversation

Request an audit

Send the basics as a prefilled email draft. No form data is stored by the site.