AI copilot, designed for teacher work
Lerarius is not a ChatGPT wrapper — it is a collection of concrete teacher tasks, each with structured input and source-grounding: every suggestion shows the exact data the AI used, so the teacher can audit the output.
What breaks today
Generic AIs ask you to write a paragraph, hope you provide enough context, and produce report phrases that sound plausible but are factually wrong. The teacher doesn't know on what data the AI based its claim.
How EduVlaanderen fixes it
Six capabilities with structured input: report phrases (per student + period), eindterm tagging (on an activity), class issue analysis (per class with time window), parent thread summary (per thread), differentiation suggestions (per assignment with class distribution), lesson plan draft (per eindterm codes + duration). Every response logs the source data in a JSONB `sources` array.
Source-grounded workflow
- Teacher picks capability + context (class ID, assignment ID, student ID, etc.).
- Backend collects structured data: e.g. for class-issue → attendance last 14 days, competency distribution, open assignments.
- Data sent to Ollama with a Lerarius-specific prompt that explicitly says: "use only these facts, no made-up names or numbers".
- Response stored with `sources` JSONB array — what the AI "saw".
- Teacher sees the output beside an expandable "Sources" section and can compare before accepting or editing.
- On accept/edit: persisted in lerarius_response.teacher_edit — feedback for future fine-tuning.
Source-grounding as professional duty
A teacher writing a report phrase about a student is professionally accountable for what it says. A report phrase from an AI without source citation cannot be professionally endorsed. By linking every Lerarius response to its source data, the teacher remains the final authority — the AI does only the typing work, not the judgement.