The method
How one session works — and why it gets sharper every time you use it.
Four moves, every session
Practice in real format
You take a short, timed set. Questions are generated to your exam's real blueprint — the section weightings, the answer-choice traps, the phrasing. Practising off-format is why so much prep doesn't transfer on test day.
Detect the weak spots
Every answer updates a per-topic mastery estimate. Miss two quadratic-systems questions and the model doesn't just mark them wrong — it lowers your mastery on that topic and raises how often it resurfaces.
Explain only the gaps
You get worked explanations for the topics actually costing you points — routed to a stronger model — and nothing for the topics you've already locked in. Your time goes where the score is.
Predict & plan
Each session ends with a predicted score and a plan for tomorrow, ordered by impact: weakest, highest-weight topics first, plus a review nudge before a shaky topic slips back.
Why it's defensible
A model of the student, not a transcript
The chat is disposable. The asset is the evolving estimate of what you know and what predicts your score — that's what makes the drilling smart.
Two-tier model routing keeps margins high
Generation and simple explanations run on a small, cheap model. Only high-value moments escalate. It reads as one smart tutor; the cost profile is nothing like running everything on a frontier model.
Compounding data advantage
Which questions predict which final scores is learned from real outcomes. Every cohort makes the predictions tighter — a moat a wrapper can't clone.
A real reason to fine-tune
Once the outcome data is rich enough, a purpose-trained scoring model beats prompting — genuine proprietary value on top of the base model.
By the last week it was basically only showing me the four things I still got wrong. That's the part no free tool did.
Run a session on your exam.
Free, ten minutes, and you'll see the predicted score for real.