Custom LLM for Perth and Western Australia
The disclosure first. Yes AI is a Melbourne consultancy: we are two hours ahead of Perth for most of the year and three once the eastern states go to daylight saving, we have no WA office and no WA staff, and if what you need is somebody at the mine site on Thursday, hire a Perth firm. What does travel across a time zone is the engineering, on-premises inference, running on hardware at your site, with no internet connection at all.
AWST, and the Two Things We Cannot Do
Every eastern-states vendor pitching you knows exactly what the time difference costs you, and almost none of them mention it. Here it is at the top of the page, along with the work we have to turn down.
We Are Two Hours Ahead. Three, Half the Year.
Perth runs on AWST. Melbourne is two hours ahead of you for most of the year and three once the eastern states go to daylight saving. That is the single biggest practical difference between engaging us and engaging a local firm, and handled badly it means your 3pm question gets answered tomorrow. So we take the inconvenience instead of giving it to you: meetings are booked in your business hours, which lands them in our afternoon and, for anything before your 11am, our evening. We do not offer you a 7am AWST slot to suit our calendar.
If You Need Someone at the Site, Hire Local
No Perth office, no WA staff, no WA-based subject matter experts, no local account manager. If your deployment needs somebody who can be at the mine on Thursday, or who picks up unscheduled at 8am AWST, a Perth firm has a structural advantage that no amount of goodwill from us overcomes. We will tell you that in the first call rather than discover it with you in month three.
We Are Not a DISP Member
Henderson and the WA defence industrial base get pitched hard on this, so let us be exact. Yes AI holds no DISP membership, no security clearances, no cleared personnel and no ITAR-compliant facility, and we cannot handle classified or export-controlled technical data. Note carefully what is being blurred elsewhere: air-gapping is a network topology, and we can genuinely build one. It is not an accreditation. A vendor letting those two ideas merge on a page is hoping you will not ask which one they actually have.
The Engineering Travels Even When We Do Not
On-premises inference does not care where the engineer sits. The model runs on hardware at your site, on your network, with no cloud dependency, which is the entire point in a state where connectivity is a genuine variable and OT segmentation is a genuine rule rather than a preference.
On-Premises Inference at Remote Sites
Pilbara, Kimberley and Goldfields operations run on connectivity that is expensive, intermittent, or both. A cloud-hosted assistant is unavailable exactly when a site most needs it, which is the worst possible failure pattern.
- Model weights and inference running on hardware physically at the site
- No internet connection required once deployed, none at all
- Hardware sized to your real concurrency and query volume, not a brochure tier
- Updates shipped on your schedule rather than pushed by a vendor
OT Segmentation Without an Exception Request
Where the operational technology network is segmented from corporate IT, as it should be, cloud AI is not a policy debate to be won. It is simply not permitted. Local inference removes the question rather than arguing it.
- Deployment inside the segment where the data already lives
- No egress path required, so no firewall exception to negotiate
- Read-only integration patterns where the OT side must not be written to
- Your security team reviews an architecture, not a vendor trust page
Maintenance and Operations Knowledge
Rotational crews mean the knowledge in one crew's heads is not reliably in the next crew's. That is a retrieval problem before it is an AI problem, and it is the most compelling one in the state.
- Maintenance history retrieval for the asset actually in front of you
- Fault diagnosis grounded in what happened the last time it happened
- Procedure and work instruction search in plain language
- The same knowledge base for every incoming rotation from day one
Inspection and Compliance Documentation
Assembling an inspection or compliance pack from field-captured input (photos, notes, readings) is high-volume administrative work sitting on top of judgements a competent human still makes.
- Draft inspection reports from field notes and photographs
- Certification and ticket expiry tracking across a workforce
- Gap-checking a compliance pack before a person signs it
- Voice notes turned into structured records
Where We Will Not Put AI
Under the Work Health and Safety Act 2020 (WA), and with industrial manslaughter provisions in force, a safety judgement needs a competent person making it. This boundary is designed in, not documented afterwards.
- No safety determination or risk judgement, in any circumstance
- No sign-off on a permit to work
- No decision about whether a job proceeds
- No clearing an asset as fit for service
Perth Head-Office Functions
The corporate side of a resources business (legal, finance, procurement, approvals) holds commercially sensitive information that is precisely the wrong thing to paste into a public chatbot.
- Contract and tender term retrieval across a portfolio
- Environmental approval condition tracking and monitoring support
- Procurement and supplier documentation search
- Board and disclosure documentation assembled for human review
How a WA Engagement Runs Across Two Hours (or Three)
Concrete commitments about the clock, and an early honest test of whether the hardware can actually go where the data is.
Booked in AWST. We Take the Early Start.
Meetings happen in your business hours, our afternoon, and our evening for anything before your 11am. The overlap window is scarce, so it is spent on decisions and walkthroughs. Anything that could be a written update is a written update, because a status meeting scheduled across a three-hour gap is an insult to both parties.
The Integration Question, Early and Honestly
Asset management and engineering platforms tend to have deep functionality and shallow integration surfaces, because the vendor has no commercial interest in making your data easy to extract. We test that in the first weeks, since the whole plan lives or dies there. Sometimes the honest answer is that your platform is the blocker and the real conversation is about the platform, not about AI.
Pilot on Hardware, on Your Network
For a site deployment that means a machine at the site running inference locally, inside the segment where the data already sits. Your controls govern it, your people operate it, and it keeps working when the link back to Perth drops, which is the only real test that matters out there.
Production, Then Get Out of the Way
Rolled out across the sites that need it, with operating documentation written so your people run it without us on a retainer. If the honest answer becomes "you need someone in Perth for this next part", you will hear it from us first.
What On-Premises Actually Means Here
Two claims, both of which your own network team can verify, and neither of which requires us to be in Western Australia.
No Model Provider to Disclose To
The models are open-weight and run on your hardware. There is no API call to an offshore vendor because the weights are on your side of the fence, which is a stronger position than any contractual assurance about somebody else's cloud.
- Open-weight models running on hardware you own
- No prompt, document or embedding leaves the site
- Verifiable by your network team rather than promised by us
- Appropriate for commercially sensitive operational information
- Not appropriate for classified or export-controlled technical data
Integration With WA Operational Systems
The AI layer connects to systems resources and energy businesses already run, where those systems permit it, which is exactly the question we test before you commit rather than after.
- Document management systems holding technical libraries
- Maintenance and asset management platforms, where licence and API allow
- Historian and process data, read-only, where OT policy permits
- An honest verdict when a platform simply will not let the data out
Related AI Solutions
Custom LLM for Mining
Capabilities built for Australian mining operations, from exploration documentation through to processing and rehabilitation records.
See mining AI solutions →On-Premises LLM Deployment
The technical detail of running the full stack on your own hardware, with no external network dependency whatsoever.
View on-premises options →Custom LLM for Energy and Utilities
Private AI for electricity and gas businesses, where operational documentation is both vast and unsearchable.
See energy solutions →Frequently Asked Questions
You are in Melbourne. How do you handle the time difference honestly?
Perth runs on AWST, which puts us two hours ahead of you for most of the year and three once the eastern states go to daylight saving. It is not a detail to gloss over. It is the biggest practical difference between working with us and working with a local firm, and most eastern vendors simply do not mention it. Our approach is to take the inconvenience rather than hand it to you. Perth meetings are booked in your business hours, which means our afternoon and, for your early slots, our evening. The overlap window is spent on decisions and workshops rather than status updates that should have been written down. There is one modest compensating advantage and we will not oversell it: work you hand over at close of business Perth time has often moved by the time you open, because we started hours ago. That is real, but it is not a reason to hire us.
Should we just hire a Perth firm instead?
Sometimes, and we will say so. If your engagement needs frequent unscheduled contact, regular site attendance, or someone who can be in a room at short notice, a Perth firm has a structural advantage and no amount of good intentions from us overcomes it. Where we are the better answer is narrower and more specific: you want the model running on your own hardware, inside your own segment, with no cloud dependency, and you want the people writing the design to be the people who will build it. That work is done on screens and in your network. It does not improve because somebody caught a flight to Perth.
Can it really run at a Pilbara site with no internet?
Yes, and this is the part of the pitch that is genuinely strong rather than merely reassuring. When a site has limited or unreliable connectivity, we deploy the model weights and the inference stack on hardware at the site itself, so queries are answered locally with no cloud round trip. Once it is deployed it does not need an internet connection to work, not a slow one, not an intermittent one, none. The same architecture is the answer where OT network segmentation means cloud connectivity is not permitted regardless of whether it is available. Hardware is sized against your actual concurrency and query volume during the assessment, and the honest constraint is cost: local inference means buying GPUs, and we would rather show you that number early than discover your appetite for it late.
Are you a DISP member? Can you do defence work at Henderson?
No, and no. Yes AI holds no membership of the Defence Industry Security Program, no security clearances, no cleared personnel and no ITAR-compliant facility. We cannot handle classified or export-controlled technical data, and we will not take work that requires it. This matters more here than the marketing suggests, because the words in this category are deliberately slippery: a vendor can truthfully say they build "air-gapped deployments", so can we, and let you infer that this satisfies DISP. It does not. The requirement runs to the supplier, its facility and its people, not just to the network diagram. If you need a DISP member, you need a different supplier.
Can AI help with safety-critical work?
It can help around it, and we are firm about the boundary. Under the Work Health and Safety Act 2020 (WA), with industrial manslaughter provisions in force, safety judgements need a competent human making them, full stop. What AI legitimately does is reduce the administrative load competing with those judgements: drafting an inspection report from field notes and photos, tracking which certifications expire when, assembling incident documentation, surfacing the maintenance history relevant to today's job. The honest framing is that this gives your safety people more time to think and gives you a better paper trail. Any vendor offering you AI that makes a safety determination should be shown the door, and if we ever propose it, show us the door too.
Which WA regulations affect a deployment like this?
The federal Privacy Act 1988 (Cth) and the Australian Privacy Principles are the baseline for private-sector businesses, and they govern personal information regardless of what processes it, including disclosure to an overseas model provider, which is a disclosure and needs treating as one. Self-hosting largely sidesteps that particular question by removing the provider. Australian Consumer Law applies to anything your AI says to a customer. WA-specific: the Work Health and Safety Act 2020 (WA) most shapes design in this market, particularly where automation touches safety-critical processes or the records evidencing them. The Privacy and Responsible Information Sharing Act 2024 (WA) introduces a privacy regime for the WA public sector and is being phased in, relevant if you hold WA government work or act as a contracted service provider. We map the ones that actually apply to your use cases rather than reciting the list. This is general information on how these obligations typically apply, not legal advice, and we will say plainly when a question needs your lawyers rather than us.
Sources and further reading
- Australian Privacy Principles (Office of the Australian Information Commissioner)
- The Privacy Act 1988 (Office of the Australian Information Commissioner)
- Consumer guarantees under the Australian Consumer Law (Australian Competition and Consumer Commission)
Two Hours Ahead, and Honest About It
One call, in your business hours, from someone who will take the early start. You will get a direct answer on whether on-premises inference is worth the hardware for your sites, including, when it is true, that you should hire a Perth firm instead.