Scientifically backed adaptive learning

We estimate mastery per topic with Bayesian Knowledge Tracing [1], one of the most widely used models of student knowledge in adaptive learning. It treats each skill as learned or unlearned and updates that probability after every attempt. We extend it two ways: attempts are scored on a continuous scale instead of right or wrong [2], and harder problems carry more evidence than easier ones [3].

Mastery is tracked separately for every topic, so a strong record in arrays never hides a gap in graphs. That gap is exactly what an interviewer finds if it goes unmeasured. A problem can usually be solved more than one way, so we credit the topics your solution actually used, not every topic the problem is tagged with. When your code is accepted, an AI reviewer reads it, identifies which topics it draws on and how heavily, and moves each topic's mastery by its share.

Each attempt's evidence is your time against an interview benchmark, any hints or wrong submissions, and whether the accepted solution hits the optimal time complexity, which the same reviewer checks against the problem's editorial. Finishing always beats giving up. More help never scores higher. Repeating a problem counts once.

Every problem in the bank is calibrated individually onto one continuous difficulty scale, so even problems at the same official level are ordered from easiest to hardest. Each topic keeps its own level on that scale, updated after every attempt by stochastic approximation [4]. Steps start large, so a handful of problems place you, then shrink once your results start alternating [5]. The same up-down logic finds perceptual thresholds in psychophysics [6].

The target is desirable difficulty: practice that is effortful but achievable, which improves long-term learning even though it feels harder in the moment [7]. Research on optimal training finds that learning is fastest when some attempts fail, not when every one succeeds [8]. Your study mode sets how hard to push, and the level you pick at sign-up is only a starting point that a few problems correct.

Spreading practice out over time improves retention compared with massing it together [9]. Problems you've solved come back for review at growing intervals, and the ones you solved cleanly wait longest.

Failures come back sooner. Making a mistake and then correcting it is a powerful way to learn [10], so a problem you gave up on returns before one you solved.

Recognizing which technique a problem calls for is often the hardest part of an interview. Flashcards train that skill directly: each card shows a problem statement, and you retrieve the approach and its complexity from memory against a timer, while an AI grader either passes the card or gives you a Socratic nudge. Retrieving an answer from memory strengthens it far more than re-reading it does [11, 12].

Cards follow a spaced schedule: missed cards come back quickly, and cards you know return at growing intervals. In a review of the evidence on common study techniques, practice testing and distributed practice were the two rated highest-utility [13].