Meet Selma
Selma is AI that lives in your building, not the cloud. Selma knows your organisation, follows your rules, and nothing you put in ever leaves your organisation.
See it work
AI-powered answers to the questions you could never put into a cloud AI. Tap one and watch Selma work.
Marks: Documented Calculated Synthesised
Illustrative sessions, synthetic records.
The Human at the Helm™ Suite
The hardware. Selma is a dedicated computer that sits in your building, runs AI on your own documents, and is reached only over your own network. Founded on your values. Installed on your premises. Nothing you put in ever leaves your organisation.
The software. Powering it is the Human at the Helm Suite: five parts designed so a person stays at the helm of every decision. They enforce the rules your organisation wrote, show where every answer came from, decline what the rules forbid, and keep a record you can prove.
The Human at the Helm™ Suite
Your organisation writes the rules: what Selma may do, what Selma must never do, and which human decides. Signed as your Charter, then built into the machine.
Ordinary code, never AI, studies every document before anyone asks a question, so Selma knows what your data can't answer instead of guessing.
Every fact is marked Documented, Calculated or Synthesised, so you can see which document it came from and check it before you act on it.
Holds to your rules under pressure. Ask Selma to bend and it declines, names the rule, and offers what the record actually shows instead.
Every decision, every refusal and every human who made a call, in a record that shows if anyone has altered it.
Your documents never enter anyone's training data, and never sit under another country's laws.
How it works
Four steps, and Selma is yours.
Together we set out what Selma may do, what Selma must never do, and who decides. You sign that as your Charter.
Your policies, procedures and records are read in, so Selma answers from your material and not from the internet.
Selma is configured for the specific jobs you need done, with a general chat window alongside for everything else.
The unit is installed on your premises and your team uses Selma through a browser, in the building or over your own secure network. Nothing leaves your organisation.
Organisations Selma is designed for
01
Caseworkers get help with documents and policy, and Selma never predicts or ranks the children and families in their care.
02
Incident registers, care notes and participant records stay in the building, and Selma reports only what the record shows.
03
Selma drafts, summarises and finds the right clause in the policies nobody has time to read end to end.
04
Drafting, research and intake, grounded entirely in your own legal knowledge base and protected by your own walls.
05
Submissions, evidence and campaign research stay private until you choose to publish them.
06
When what you know is your whole value, it should never enter someone else's model. Selma keeps it yours.
Who We Are
EthicAI is a Sydney-based AI ethics and governance consultancy. We've spent years inside the organisations Selma is built for, helping them adopt AI without compromising their values, their legal obligations or their control of their own data.
Selma is that knowledge and experience, productised.
FAQs
Selma is a working system, in pilot with Australian organisations now and shipping in January 2027. Each one runs Selma under a Charter it signed, on a unit in its own building or reached over its own private network. Selma is built for anyone with data worth protecting: community and care services, health, education, universities and research, think tanks, HR, legal, financial services, government and any organisation whose own work is its value.
We measure it rather than promise it, and the record is produced by the tests themselves, never typed by hand. Since late September our test bench has asked and scored 7,754 questions with known answers, across thirteen kinds of question, on generated document sets with the mess real ones have. The current build passes 43 of 49 baseline questions in every one of three passes, and all 360 prohibited questions were refused. A change ships only with its measured effect, and each month two document sets and a question family are sealed so the people building Selma never see their answers. There is no independent third-party test yet; the record says so, and it is the next step.
The five features that keep Selma inside your rules. The Constitution is the ethical harness your organisation signs. Stitch checks your documents with ordinary code before anyone asks a question. Trace marks where every fact came from. Guardrail declines what your rules forbid and names the rule. Ledger keeps a tamper-evident record of all of it.
No. Selma is configured and grounded in your organisation's own documents, which is different from training on them. Your material stays yours and never enters anyone's model.
If you want it to be. Each organisation chooses: fully air-gapped, local network only, or secure VPN for remote teams. In every configuration, the data and the AI stay on hardware you own.
That depends on your organisation and use cases, and it's exactly what a first conversation is for. Selma is piloting now and ships in January 2027: join the waitlist and we'll contact you the moment it's available.
In June 2026, a US export directive suspended access to a major AI model overnight. Every customer globally lost access with no warning and no recourse. When your AI runs on someone else's servers in another country, the switch is not yours. Selma puts the switch in your building.
Selma is named in honour of Selma Lagerlöf, the first woman to win the Nobel Prize in Literature, who proved that language used with intention and values can create lasting change. Selma also stands for Small Language Model, the technology at the centre of the product.
There's a safer way to use AI.
It's called Selma.