Ephraim

Ephraim is the AI assistant for Lessing-Gymnasium Karlsruhe. It runs on a single machine inside the school building — 600 students and teachers, zero cloud.

Privacy first

Ephraim is used by minors. That's the first thing that shaped every technical decision.

When you sign up, the system generates a unique encryption key just for you. That key locks your conversations so that nobody else can read them — not other users, not teachers, not whoever runs the system. The key itself is sealed behind your password in a way that makes it practically impossible to recover without it. Even with full access to the database, all anyone would find is noise.

The key only exists in memory while you're logged in. When you log out, it's gone. Nothing readable ever touches the disk.

This isn't just a policy — it's enforced at the level of the data itself, where it can't be worked around.

The hardware

Everything runs on an NVIDIA DGX Spark — a machine roughly the size of a hardback book, weighing 1.2 kg. It has 128 GB of unified memory, which is enough to hold a 120-billion-parameter language model alongside everything else Ephraim needs, without any slowdown. The entire school's AI fits on one desk and draws about as much power as a gaming laptop.

The model

Ephraim uses a fine-tuned version of an open-source 120-billion-parameter model, adapted specifically for the school. It knows the curriculum, internal structure, and how the school operates. Students and teachers interact through a familiar chat interface, but the model behind it actually knows their context.

Memory

Ephraim remembers things across conversations. Not in a vague way — it builds a real profile for each user over time. If you mention you prefer shorter explanations, or that you're working on a specific topic, Ephraim notes it. A background process periodically reviews and organises these notes into a coherent picture that shapes every future conversation.

Beyond profiles, completed conversations are stored and indexed by meaning. When a new question relates to something discussed weeks ago, Ephraim finds it and brings the relevant parts into the current context — without you needing to reference it yourself.

Teacher projects

Teachers can build structured projects for their classes. A project gives Ephraim a specific role — a tutor for a particular chapter, a writing assistant for a specific assignment. Teachers attach the relevant material, set the scope, and optionally add prompts to guide students in.

Students join with a code. Inside, they get an assistant that knows the material and stays on topic. Teachers can see who's participating, but not what anyone said.

Streaming and resilience

Responses stream back word by word as the model generates them. Each fragment is encrypted before it passes through the system's internal buffer. If a student's connection drops mid-response — laptop closed, Wi-Fi cut out — the stream holds. When they reconnect, it resumes exactly where it stopped.

A separate process handles updates and restarts independently from the rest of the system, so changes can be pushed from a dashboard without taking anything down.