The school’s AI stays in the school.
A full assistant for ~800 students and teachers, running on one machine inside the building. No vendor API. No prompt ever leaves the premises.
Classroom AI keeps failing the same check.
Every prompt is a school record.
A question typed by a fifteen-year-old is a school record the moment it exists. Send it to a vendor API and it has left the building, the jurisdiction, and your control — before anyone has read it.
Consent nobody can give.
Minors cannot meaningfully consent to third-party processing, and a school cannot consent on their behalf. Most classroom AI stops at exactly this sentence.
The thing you approved is gone.
Providers deprecate, re-tune and re-route. Whatever a data-protection officer signed off in spring is not necessarily what answers a student in autumn.
A meter on curiosity.
Per-token billing means the cheapest classroom is the one that asks least. That is precisely backwards.
Everything local.
Nothing leaves.
One box on a shelf answers every question the school asks it. The interesting engineering here is not the model — it is that nobody, including the person holding the server, can read what anyone typed.
Three decisions, and the rest follows.
One machine, in the building.
An NVIDIA DGX Spark — 128 GB of unified memory, about the size of a hardback book, drawing roughly what a gaming laptop does. It sits on a shelf in the school. That is the entire deployment.
An open 120B model, adapted without retraining.
The model is not fine-tuned on student work. It is shaped by everything around it: retrieval over curated school material, role and policy layers, and prompts that know the timetable, the structure and the house rules.
Encryption that survives the server admin.
Each account generates its own key, sealed behind the user's password. It exists in memory while they are signed in and nowhere else. With full database access you get noise — that is the design, not the policy.
The questions that actually get asked.
01Is this running for real students?+
It ran a controlled pilot through July 2026. It is not generally rolled out, and it will not be until an external security review, threat modelling with professionals, the teachers' conference and the data-protection review all clear it. A system used by minors does not get to skip that.
02Which model is it?+
An open-weights 120-billion-parameter model, running locally. The point is that it is open and it is ours — nothing about the deployment depends on a vendor keeping an endpoint alive or a price stable.
03Can teachers read student conversations?+
No. Not teachers, not administrators, not whoever runs the server. Conversations are encrypted under a key derived from the user's own password. Teachers running a class project can see who joined, not what anyone typed.
04What happens when the connection drops?+
Responses stream token by token, and each fragment is encrypted before it touches the internal buffer. Close the laptop mid-answer and the stream holds; reconnect and it resumes where it stopped.
05Does it remember me between conversations?+
Yes, deliberately. It builds a profile over time — how you like explanations, what you are working on — and can pull in a relevant conversation from weeks ago without you citing it. All of it under the same per-user encryption.
06Can another school run this?+
Not yet, and honestly not soon. It is built for one school's structure and one school's governance. Talk to me if you want to compare notes.
In pilot. On purpose, slowly.
An external security review, threat modelling, the teachers’ conference and the data-protection review decide when this scales. I would rather list what is still open than pretend it is not.
No signup. No demo instance. It runs in one building.