Thynkr — adaptive learning for Australian Year 11 and 12 students.

A Year 12 student in Queensland is preparing for the external exam that will decide their ATAR, which will decide their degree, which will decide the first decade of their working life. The tools available to them are a thick textbook, a teacher managing thirty others, and a generic chatbot that forgets them between sessions and cannot tell whether they are stuck on the chemistry or stuck on the algebra inside the chemistry. They do not know what they do not know. The misconception they carried in from Year 10 is still sitting there, invisible, producing wrong answers that look like careless mistakes. When the judgement about what to study next leaves the student's head and lands in a textbook's table of contents, the practice stops being learning and starts being ritual.

Every attempt, chat turn, hint request, and skipped question is an event against the student's profile. A knowledge tracer updates per-concept mastery, a misconception detector tags the error against the syllabus's known failure modes, and a behavioural collector watches for the softer signals — how long the student sat on the question, whether they asked for a hint before trying, whether the explanation landed. The question selector reads from that projection, not from a fixed sequence, and proposes the next item. The AI tutor proposes an explanation in the student's style profile. The student decides whether it worked.

The loop is deliberately student-in-the-middle. The tutor never marks an answer "wrong" and never awards points for getting one right — it asks the next Socratic question and lets the student resolve it. The career, subject-selection, and mentor agents sit on the same event store and the same memory layer, so advice about which subjects to take in Year 11 is written from the same understanding of the student that drives the next chemistry question.

• Events over CRUD — mastery, misconceptions, and study style are not fields on a student row; they are projections over a long stream of practice attempts and conversations, which is the only shape that survives a student changing, a syllabus changing, or a new subject being added.
• Three-tier context — curriculum (the QCE syllabus graph and RAG corpus), student memory (durable facts the mentor has learned), and session (the current conversation) are composed per call, so the tutor sees the same student the career advisor sees without either service owning the other's data.
• Skills over controllers — tutor, mentor, career, subject-selector, ATAR optimiser, and UCAT prep are separate agent skills over a shared learner model, which is what lets a new subject or a new advice surface ship without touching the others.
• Feedback as interaction — a wrong answer, a favourited response, a skipped hint, and a request for a similar question are all first-class signals that reshape the next selection, rather than telemetry collected for a dashboard nobody reads.