Seven practices, one engineering team. We design, build, benchmark and operate the compute layer beneath your hardest problems — on-premise, hybrid or sovereign cloud.
A genomics pipeline and a trading engine are not the same machine. We start from your regulatory perimeter, your data gravity and your deadlines — then design backwards.
Industries · Government & Public Sector
Three questions decide the design of a government platform before any component is chosen. Where does the data physically sit. Who can be legally compelled to hand it over. What evidence will the assessor and the authorising officer ask to see. Everything downstream is a consequence of those answers.
The constraint we design around first
Most government infrastructure projects fail their security milestone for documentary reasons, not technical ones. The control was implemented and never written down, or written down and never tested. We design the evidence trail at the same time as the platform.
A delegate inside your agency signs the authority to operate. That person carries the residual risk personally. Their question is narrow: what is the exposure, what compensates for it, and who verified that. We build the answer as a by-product of the build rather than a retrofit at the end.
We would rather hand an agency a short list of accepted weaknesses than a long claim of total compliance. The first survives scrutiny.
Data sovereignty, unbundled
A platform can satisfy the first and fail the other two, and the marketing material will still say sovereign. Ask which of the three you are buying, and ask for the mechanism rather than the adjective.
Where the bytes rest and where they are processed. The easiest of the three to verify and the least meaningful on its own. Check the whole surface: primary storage, replicas, backups, search indexes, message queues, crash dumps, observability pipelines and any managed service that ships metadata offshore.
Test — name the region of every persistent store and log sink
Who can technically reach the data, from where, and under whose supervision. In-country storage administered by an offshore follow-the-sun support team is not operationally sovereign. This is a question about privileged access paths, break-glass procedures, personnel screening and support hours.
Test — list every human and service account with a path to plaintext
Which legal systems can compel disclosure from the operator, regardless of where the hardware is. A provider incorporated offshore, or controlled by an offshore parent, can be subject to foreign production orders over data held in Australia. Bilateral arrangements change the detail; corporate structure determines the exposure.
Test — trace ownership to the ultimate controlling entity
| Ask this | Property under test | What a weak answer looks like |
|---|---|---|
| Name every region in which a copy of this data exists, including backups older than 30 days. | Residency | "All data is stored in Australia" with no statement about backup or disaster-recovery targets. |
| Who holds the encryption keys, and can the operator decrypt without our participation. | Operational | Provider-managed keys described as customer-controlled because the customer can rotate them. |
| From which countries can support staff open a privileged session, and is that enforced technically. | Operational | A policy commitment with no network or identity control behind it. |
| Which entity signs the contract, and who ultimately controls that entity. | Jurisdictional | An Australian subsidiary signing while the parent retains technical administration. |
| What is the process if a foreign authority serves a production order on the operator. | Jurisdictional | No documented notification path to the agency, or a contractual gag clause. |
| Where do telemetry, crash reports and support attachments terminate. | Residency and operational | Silence, or an exclusion for "diagnostic data" buried in a sub-processor list. |
Assurance
Described here at a descriptive level so procurement and engineering teams share one vocabulary. Each framework is revised; check the current published version before you rely on any summary, including this one.
The Infosec Registered Assessors Program registers individual assessors to evaluate a system against the Information Security Manual. The output is an assessment report describing how controls are implemented and where they fall short. It is a statement of findings.
It is not a pass mark, a licence, or a transferable badge. No assessor grants authority to operate. That decision belongs to an authorising officer inside the consuming agency, who weighs the findings against their own risk appetite and the sensitivity of the data. Two agencies can read the same report and reach opposite conclusions, and both can be correct.
Practically, this shapes engineering. A narrow, well-documented boundary assesses faster than a sprawling one, so we push cleanly separable components out of scope early — identity, egress, data science sandboxes — rather than arguing about them during the assessment.
The Information Security Manual is published by the Australian Signals Directorate as a risk framework, not a compliance checklist. Controls are expressed with applicability by classification, and agencies apply them through a risk-management process rather than tick by tick.
We treat it as a specification with real engineering consequences. Several control families drive topology directly, and discovering that late is expensive.
Classification handling under the Protective Security Policy Framework sits alongside it, and the two must agree. If a system carries PROTECTED data, the marking rules, the aggregation risk and the physical zone of the facility are part of the same design problem as the storage tier.
Aggregation deserves separate attention. Datasets that are individually OFFICIAL can become PROTECTED once joined, because the combination identifies people or reveals capability. We ask about intended joins during design, not after the data lake exists.
Eight mitigation strategies, each with defined maturity levels. Most agencies already know the list. The difficulty is that several of the eight are only sustainable if the platform underneath was built with them in mind.
01
Application and operating-system patching to a maturity target means a tested pipeline, not a maintenance weekend. Immutable images, staged rings and a rollback path make aggressive cadence survivable. Without those, teams choose between the control and availability.
02
Allow-listing executables is straightforward on a fleet of identical nodes and painful on hand-built servers. Configuration as code and golden images are the difference between maturity two and a permanent exception register.
03
Just-in-time elevation, separate administrative credentials and session recording need an identity design that anticipated them. Retrofitting privileged access onto a flat estate is one of the more common causes of a failed re-assessment.
04
The control is about recovery, not copies. We size for a tested restore inside the stated recovery objective, keep at least one immutable copy outside the production trust boundary, and schedule restore drills as a calendar obligation.
Maturity requirements for Commonwealth entities are set through policy that changes over time. Confirm the current obligation for your entity type rather than inheriting a target from an older programme.
Commonwealth hosting policy has, in recent years, required certain workloads to sit in facilities certified under a hosting framework, with conditions on ownership and control of the data centre operator. The framework has levels; the level required depends on the workload and the policy in force.
Treat the framework as one input among several. Certification of a facility says nothing about the tenant's own controls, and it is not a substitute for an assessment of the system. We check four things when selecting a site for government work.
Density is where facility choice and workload collide. GPU racks for government AI work draw far more than a legacy hall was built for, and the certified facility with the right ownership profile is not always the one with the cooling capacity. That trade-off should be surfaced in the business case, not discovered during installation.
Citizen-facing infrastructure
A lodgement deadline, a census night, a disaster declaration, a payment indexation. Government demand is not a smooth distribution; it is a flat line with cliffs. The engineering question is what the platform does in the first ninety seconds of a cliff.
Every system has a breaking load. What separates a bad day from a public inquiry is whether the failure sequence was designed. We decide in advance which functions shed first, and make sure the citizen never loses a submission they believed had succeeded.
Then we test it. Load rehearsal against a production-shaped environment, at a multiple of the forecast peak, with the failure modes deliberately triggered. A rehearsal that only tests the happy path tells you almost nothing.
Procurement
Engineers propose architectures that procurement cannot lawfully buy, and procurement runs processes that cannot describe the architecture. We work on both sides of that gap because delay is usually procedural, not technical.
Whole-of-government and state panel arrangements cover most infrastructure and ICT services, each with its own scope, categories and rules about what can be bought through it. A design that sits outside every available arrangement will need an open approach or a documented exemption, and that changes the timetable by months. We check route and scope first.
Under the Commonwealth Procurement Rules value for money is the core rule, and it is assessed on whole-of-life cost against outcome rather than lowest price. For infrastructure that means a costing that includes power, cooling, facility, refresh, exit and the staff hours to operate it. We provide the model in a form your evaluators can interrogate, including the assumptions we are least confident about.
The Indigenous Procurement Policy sets targets by volume and value, allows direct engagement with Indigenous enterprises under a specific exemption in the procurement rules, and imposes mandatory minimum participation requirements on contracts above defined thresholds in certain sectors. It is a real obligation with reporting behind it, not a statement in a tender response. Structure the supply chain to meet it, and be able to show how.
Transition-out terms, data extraction formats, egress cost and the disposition of custom configuration all belong in the initial agreement. Agencies that leave exit to a later negotiation pay for it. We write the exit plan as a deliverable and price the extraction once, at the start.
Commonwealth and state records legislation governs how long information is kept and when it must be destroyed or transferred. Retention is therefore a storage-tier decision and a deletion-capability decision at once. A platform that cannot demonstrably destroy a record on schedule creates a compliance problem that no amount of capacity solves.
AI with guardrails
Australian government AI policy has moved toward accountable officials, published transparency statements and assurance before deployment. The infrastructure has to make that provable, which is mostly a data-governance problem rather than a model problem.
Private inference inside the agency's perimeter is the starting point, not the finish. The controls that matter are the ones that let an agency answer, two years later, which data trained this and who saw that output.
Aggregation returns here. A retrieval corpus assembled from several OFFICIAL sources can carry a higher effective classification than any of its parts, and the model will happily join them. We classify the corpus, not just the documents.
The hard part of government AI is not serving the model. It is being able to prove, later and under scrutiny, exactly what went into it.
Questions we are asked in evaluation
If an answer here contradicts something a salesperson told you, the answer here is the one we will defend in a tender clarification.
No vendor can hand you an authority to operate, because that authority is issued inside your agency. What can be reused is assessment work over a defined platform boundary, which shortens your own process. Ask any vendor for the assessment date, the ISM version used, the assessed boundary and the change history since. If those four things are not available, the claim is not usable in your risk decision.
No. Hosting location covers residency only. Operational sovereignty asks who can technically reach the data and from where. Jurisdictional control asks which legal systems can compel the operator to disclose it. A platform can be entirely Australian-hosted, administered from overseas, and owned by a foreign parent. All three properties need separate answers, which is why the comparison table above exists.
Not at all, and we will often recommend keeping it. Public cloud is the right answer for elastic, lower-classification and citizen-facing web workloads. The cases where dedicated sovereign infrastructure wins are narrower: steady high-utilisation compute where consumption pricing is expensive, workloads with strict jurisdictional constraints, and estates where egress charges distort architecture. We will tell you which of your workloads is which, including the ones that should stay where they are.
Usually not, and over-classifying is expensive in both money and usability. The question worth asking is whether aggregation pushes the effective classification up. Joining several OFFICIAL: Sensitive datasets can produce a holding that warrants PROTECTED handling. We look at the intended joins and the analytical use before recommending a level, because the decision is yours to make and it should be made on the aggregate.
By deciding the degradation order in advance and rehearsing it. Admission control holds the front door, durable intake protects submissions already accepted, and non-critical functions are shed in a documented sequence. We agree that sequence with the service owner, because which function is sacrificed is a policy judgement rather than an engineering one.
Only where that client has given written consent, and named references only where the contract permits it. We would rather run a benchmark on your own workload than show you a number from someone else's. Figures quoted without a stated measurement method and a consenting source should be treated as marketing by any evaluator.
Let's Talk
Bring a real constraint — a control you keep failing, a lodgement window that breaks the platform, a sovereignty requirement nobody can define. You get an engineer's read on feasibility and cost, and a clear statement of what we cannot evidence.
Sydney · Melbourne · Canberra. Australian owned, Australian staffed.