Skip to content
Talk to an engineer

Sovereign AI, built in Australia.

Sight is computer vision built in Melbourne, trained and run in Australia, and measured in the open. This page says exactly where each part runs, what is yours to choose, and who we want to build the rest with.

Talk to an engineer Read the ledger Work with us

Sovereign is a word anyone can print. Here is what we mean by it.

  • 01

    Built here

    Designed, written and supported by an Australian company, by engineers you can meet.

  • 02

    Run here

    Your photos, your models and your results stored and processed inside Australia.

  • 03

    Answerable here

    A contract under Australian law, and a straight answer to “where does this part run, and who operates it?”

  • 04

    Measured here

    Claims tested on real Australian field photos and published, including the ones that failed.

Where your photos, models and results go.

The AI, and the data it works on, part by part. 8 of 9 parts run in Australia. The other, reading text, is yours to choose, because most frontier language models are run overseas.

  • The company and its engineers

    In Australia

    Melbourne

    Raku Technologies Pty Ltd is an Australian company. Sight is designed, written and supported from Melbourne, and your contract is with an Australian company under Australian law.

  • Your photos and trained models

    In Australia

    Stored in Australia

    Every job photo, every label and every model version trained on them.

  • Projects, labels and results

    In Australia

    Database in Sydney

    The record of what was checked, what the answer was and why.

  • Finding and counting

    In Australia

    Runs in Australia

    The detection model trained on your photos runs on machines in Australia. A photo that only needs parts found and counted is processed here from start to finish.

  • Training

    In Australia

    Runs in Australia

    Models are trained on GPU machines in Australia, and scored there on photos they have never seen.

  • The rules that judge a job

    In Australia

    Runs in Australia

    The check engine makes no network call at all. It was written that way so the same rules can later run on a phone with no signal.

  • The data centres

    In Australia

    In Australia

    The machines and storage Sight runs on sit in data centres in Australia. We would like them to be Australian-owned as well: if you run one, see the initiatives below.

  • Reading text

    You choose

    Australia or overseas

    Some checks need text read from the photo: a label, a serial number, a reading. That step is optional, and where it happens is your choice, because most frontier language models are run overseas. Choose overseas for the widest choice of models. Choose Australia: send the reading your own device already made, and Sight uses it without sending the photo anywhere (live); running on a model hosted here or on your own language-model account is rolling out. Or leave reading off, and nothing leaves Australia.

  • The detection model

    In Australia

    Open-source design, run here

    Sight’s detection models use an open-source design that was developed overseas. The copies trained on your photos are trained, stored and run in Australia.

Checked 3 October 2026. This is our own statement, not a certification: Sight has not been through a government security assessment. If you need one of these parts to run somewhere else, ask an engineer.

Bringing the frontier home.

Most frontier language models are run overseas. This is the work that lets every part of a check run in Australia, each item marked with how real it is.

  • Live

    Leave the reading step off

    A check built only from finding, counting and position never sends anything overseas. For many jobs that is the whole check: is the part there, is it the right one, is it in the right place.

  • Live

    Readings made on the device

    On 44 real install photos, a phone’s own text recognition read serial numbers in 0.2 seconds with one error, with no signal and nothing leaving the phone. Send that reading and Sight uses it instead of the photo; the check answers at once and nothing goes overseas.

  • Rolling out

    Your own language-model account

    Connect the language-model account your organisation already has, and reading runs under your contract, in the region you chose. Or point a check at a model hosted in Australia. Sight runs the rules and records who read what, so every verdict can still be audited.

  • Rolling out

    Your storage, your region, your cloud

    Photos read in place from storage you own, with processing in the region you choose or inside your own cloud account.

  • Rolling out

    Small models of our own, run locally

    Small reading and decision models that run on our machines in Australia or on the device itself, so the common questions never need a large overseas model at all. We have measured several open ones on real field photos; none is good enough yet without training on your own examples.

  • In research

    One engine that needs no data centre

    The check engine rebuilt to run identically on a server, in a browser and on a phone. The most sovereign place to process a photo is the device that took it.

Research done here, on real work.

Every number below was measured by Raku on real field photos, and none is a vendor benchmark. We publish what failed beside what worked.

This is our own lab’s work. We would like Australian research partners: see the initiatives below.

Why this is a national question.

What Australian governments have said and funded, each with its source.

2 Dec 2025 the National AI Plan was published

The national plan commits to sovereign capability, and never says what it is.

The ministers’ foreword to Australia’s National AI Plan commits to “building sovereign capability in AI”. The plan does not define the term. In practice, across government, it has come to mean “hosted in Australia”.

The same plan says where Australia’s edge lies: in targeted, high-value AI products for particular sectors. Checking real work in the field is one of those.

Source: National AI Plan, Department of Industry, Science and Resources

Nov 2024 Recommendation 4 of the Senate committee on adopting AI

The Senate asked for more support for sovereign AI capability.

The Senate Select Committee on Adopting Artificial Intelligence recommended that the government keep increasing its support for sovereign AI capability, focused on the areas where Australia already has an advantage.

Source: Senate Select Committee on Adopting AI, list of recommendations

17 Nov 2025 federal procurement rules began to define an “Australian business”

Federal buying rules now favour Australian businesses for smaller contracts.

The Commonwealth Procurement Rules define an Australian business by ownership, tax residency and where it is run from. For contracts from $10,000 up to the procurement threshold, agencies are told to invite only Australian businesses, unless an official records why that is not appropriate.

It is the strongest lever government has today for buying Australian technology, and it says nothing yet about where an AI system does its thinking.

Source: Commonwealth Procurement Rules, Department of Finance

Expectation 1 for data centres and AI infrastructure, March 2026

Data sovereignty is now part of the national interest test for AI infrastructure.

The government’s expectations of data-centre and AI infrastructure developers begin with the national interest, which they say includes national security and data sovereignty. They are expectations used to prioritise approvals, not law.

Source: Expectations of data centres and AI infrastructure developers, DISR

1 July 2026 the critical technologies list was last updated

AI and advanced imaging are both on the list of critical technologies.

The List of Critical Technologies in the National Interest covers fields where Australia needs dependable access or must keep its own capability. It names AI technologies, and separately advanced imaging and advanced sensors: the two halves of computer vision.

Source: List of Critical Technologies in the National Interest, DISR

The other view from the Productivity Commission, February 2024

The sceptics have a point, and it favours small models.

The Productivity Commission doubts that government sponsorship of parts of the AI value chain pays off, and argues that Australia’s advantage is unlikely to be in work that needs extreme quantities of data and money.

It says the advantage could instead be in models trained on smaller, high-quality datasets and adapted to specific uses with local data. That is a description of what Sight is.

Source: Making the most of the AI opportunity, Productivity Commission

What we think Australia should do.

Our view, as a company that builds this here. Not a statement of anyone’s policy.

  • Define “sovereign” before buying it

    The national plan uses the word without a definition, so today it means whatever the seller wants. A short public standard would do: where the data is stored, where it is processed, who operates the system, and who owns the company.

  • Ask where the thinking happens

    Where data is stored is half the question. The other half is where it is processed, and by whose model. Any supplier of AI to government or critical infrastructure should publish both, part by part, the way this page does.

  • Test AI on work it has never seen

    A model scored on its own training data looks excellent and proves nothing. Before AI is trusted to judge real work, it should be scored on examples held back from it, and the score published.

  • Fund adoption, not only invention

    As far as we could find, no federal program helps a company pay to adopt AI; that support sits in a few state manufacturing programs. The productivity gain is in mid-sized operators using AI on real work, and support should reach them.

  • Back small models that run on the device

    Sovereignty does not have to mean building the largest models. A small model on a phone in a pit, sending nothing anywhere, is sovereign by construction, and Australia can build those now.

  • Prefer Australian-run AI for critical infrastructure, or say why not

    When public money builds networks, roads and energy assets, the systems that check that work should be ones Australia can inspect, direct and keep running. Where an overseas system is chosen instead, the reason should be on the record.

Work with us on sovereign AI.

Australia will not get sovereign AI from one company. These are the people we want to build it with. None of these is a partnership yet; each is an open invitation.

  • Australian-owned data centres

    Sight runs in data centres in Australia. We would like to run it on Australian-owned infrastructure, and to offer that to customers who require it. If you operate one, we want to hear from you.

  • Frontier models, hosted here

    Most frontier language models are run overseas. If you build or host capable models in Australia, we would like to make yours the reader behind Sight’s checks.

  • Universities and research groups

    Our lab measures computer vision on real work and publishes what fails. We are looking for Australian vision and machine-learning groups to work on small models, on-device reading and honest evaluation.

  • Research companies

    If you are building a piece of the same puzzle in Australia, such as compute, models, data or evaluation, there is more to gain from building it together than from each of us buying it overseas.

  • Government and infrastructure owners

    If you are responsible for work on public assets and want it checked by systems Australia can inspect and direct, we would like to run a pilot with you.

Register your interest

Or write to [email protected] with “Sovereign AI initiatives” in the subject.

Ask us where each part runs.

For government, utilities and the owners of critical infrastructure. Bring your requirements for where data may be stored and processed. An engineer will tell you which parts meet them today, which do not, and what it would take.

By sending this, you agree to our Privacy Policy.

We store your details to contact you about the consult and keep chat records to improve our agent. We never sell your data. See our Privacy Policy.