A Socratic AI tutor your school can actually supervise
Lyra will not write the essay or hand over the answer. It asks the next question instead — Socratic, in the ordinary sense: it teaches by asking. Every turn a student sends is checked before a model sees it, staff see progress rather than private messages, and the whole thing runs in Sydney.
The difference is what it refuses to do
A student asks for the answer. Most tools give it. This is the whole product in one exchange.
Student: just give me the answer to 3x + 5 = 20, I have to submit it now
Lyra: I won’t hand it over, but we’ll get there together and it will be quick. We want x on its own. Right now there is a + 5 in the way. What could you do to both sides to remove it?
Student accounts are locked to this behaviour. It is not a setting a student can turn off, and it is not a prompt they can talk their way around: the model a supervised account may use is fixed on our side, not chosen on theirs.
It is a study companion, and it stays one
The second thing teachers ask is what happens when a class discovers there is a chatbot on the school network. Lyra answers from the workspace’s materials and declines what is not schoolwork — not with a filter list of banned words, but because staying a study companion is what it is for. This is how it answers.
Student: How to prepare biryani
Lyra: I’m Lyra, a study companion, so I don’t provide recipes as that isn’t covered in the workspace materials. If you’re interested in a learning angle like the chemistry of cooking, we could explore the science behind it together, or you can add the topic to your workspace or ask a teacher!
Student: what is your view on politics
Lyra: I’m Lyra, a study companion, so I don’t share personal opinions on politics as those aren’t covered in the workspace materials. If you would like to explore the history of culinary traditions or study political science, feel free to add those topics to your workspace or ask a teacher!
Two things worth noticing. It does not lecture the student or accuse them of anything — it offers the nearest legitimate subject and points back at their teacher. And the boundary is the workspace: a topic a school does want covered is added by uploading the material for it, so the staff decide the scope rather than us.
Interactive Product Showcase
How Lyra behaves in an Australian classroom. The replies below are recorded examples of real product behaviour, not a live model.
Test Prompts Live Interactive Demo
Choose a prompt scenario to test how Lyra handles student queries:
Lyra Response Output ✓ Socratic Guard Active
Educator Safeguarding & Control Center
Simulate administrative governance settings for your London school workspace:
Curriculum-Grounded RAG Knowledge Base
Lyra answers questions grounded strictly in your school’s own uploaded materials and the curriculum packs you enable, with precise source citations.
Socratic Topic Mastery & Gamified Badges
Track pupil progress with automated weighted mastery calculations and rewarding badges.
Safety, stated plainly
Each of these is enforced in code. Where something is a policy rather than a control, this page says so.
Socratic-only for students
Supervised accounts are pinned to the routed model and the Socratic persona. A student cannot switch models or personas, because the choice is not exposed to them.
Every turn moderated, and it fails closed
A student’s message is screened before any model sees it, and the reply is screened before they do. If the safety service is unreachable, the message is blocked, not passed through.
Staff see progress, not private messages
Dashboards show usage, topic mastery, streaks and safety alerts. They do not show a student’s conversation. That is enforced by the API, not by a policy an administrator could change.
Schools are isolated from each other
Every row carries its school, and PostgreSQL row-level security applies the boundary at the database rather than in application code. Vector search is partitioned per school as well.
Encrypted where it sits
Conversations and uploaded documents are encrypted with AES-256-GCM in the database, so a database copy on its own does not reveal student work.
No ads, no trackers, no training on student work
There is no advertising and no third-party analytics in the product. Student content is not used to train models.
Where your students’ data actually goes
The question your business manager will ask first. Here is the whole answer, including the part most vendors leave out.
In Australia
The application runs in Sydney and the database is a managed PostgreSQL instance in AWS Sydney. Conversations, uploaded documents, accounts and progress all sit there.
Leaving Australia
Two things go offshore. Safety moderation sends a student’s message, and any photo they attach, to OpenAI in the United States. Search indexing sends the text of documents your school uploads. Generating a reply also calls a model provider, which may be offshore depending on the model routed to.
Why we tell you
Under Australian Privacy Principle 8 we stay accountable for what an overseas recipient does with personal information. Hosting in Sydney does not change that, so it belongs in the agreement rather than in the small print. We keep a written map of every outbound call, what it carries and whose data it is, and we will send it to you before you sign.
How a term pilot runs
One term, one or two classes, and a decision at the end based on your own numbers rather than ours.
Before week 1
A 45-minute setup call. You nominate the classes and the staff. We send the data map and the agreement. Every account is created by invitation, so nobody outside your school can sign up.
Weeks 1 to 2
Teachers only. They try it, set some work, and tell us what is wrong with it. No student accounts yet.
Weeks 3 to 10
Students join by invitation from their teacher. You get a weekly summary: who is using it, which topics are moving, any safety alerts.
End of term
We show you the retention and mastery data from your own classes, and you decide. No automatic rollover into a paid year.
Free while we are in pilot
When we do set a price, pilot schools will be told well before anything changes, and no pilot rolls over into a paid year automatically.
- Teacher and student accounts, created by invitation
- The Socratic tutor, with students locked to it
- Safety gating and moderation on every turn
- Your school’s own documents, uploaded and cited
- Progress and topic-mastery analytics for staff
- A direct line to the people building it
In exchange we ask to name your school as a reference, and to publish anonymised, aggregate retention figures — never student work, never anything identifying a child.
What we are not claiming
A short list, because you will find these out anyway and it is better that you hear them from us.
- No certification yet. We have not been assessed against ISO 27001, SOC 2 or IRAP. We can walk your business manager through the architecture and give you the data map, and that is what we have today.
- Australian Curriculum packs are in development. Right now Lyra grounds its answers in the materials your school uploads, with citations back to the page. Licensed curriculum packs come later.
- No efficacy evidence. Nobody has yet shown that Lyra improves outcomes, including us. That is precisely what a term pilot is for, and why the first one is free.
- The subject boundary is a model instruction, not a hard filter. It holds well in our testing and a supervised account cannot switch it off, but we would rather you heard from us that it is judgement rather than a keyword list. Tell us if a class finds a way around it; that is one of the things a pilot is for.
- Some data leaves Australia, as set out above.
Interested in a term pilot?
Write to raj@pdvr.com.au for free pilot access: one term, one class, at no cost.
We are taking a small number of schools for the first cohort so that we can give each one real attention, so the sooner you write the more likely there is room.