Shaipe — Making AI-Powered LSAT Preparation Actionable

Context

The LSAT is a gating exam, and the standard route to a good score is private tutoring at $200 an hour. Several members of the RIIID team had taken the test themselves and had a clear view of the problem: the people who most need help are the ones least likely to afford it. RIIID's broader bet, across the company, was that AI could do something useful with that gap. Not replace tutors, but build a product that could carry the routine parts of preparation for students who weren't going to get a tutor either way.

Shaipe was the product expression of that bet: an LSAT study companion built around AI-driven personalisation. Predicted scores, scheduled practice, customisable preferences, and a review system anchored on bookmarked questions.

The problem

LSAT prep isn't really a content problem. The content is well understood and widely available. The actual problem is rhythm and self-trust. Students prepping for the test face four questions every time they open a study app:

Where do I stand right now?

What should I work on next?

How do I know my practice is actually moving me toward my target score?

What do I do with the questions I got wrong?

The design problem was turning AI-generated diagnostics into answers a student could act on without having to interpret a probability score.

My role

I worked on the student-facing flows. The team was four product designers under the Head of Design, with each of us owning a slice of the experience. My slice covered the predicted score surface, the scheduled-session flow, study preferences, the bookmark and review loop, and the navigation that tied those together.

Predicted score

This was the load-bearing piece. The model predicted a score based on practice performance, and the more practice a student logged, the more accurate that prediction got. Two design problems sat behind that.

The first was confidence. A predicted score with no indication of how reliable it is feels like a guess. We surfaced a confidence band alongside the headline number, and treated the band as the primary signal rather than an asterisk. Early on the band is wide, and you can watch it tighten as you put work in. That movement carries more meaning than the number itself.

The second was action. A predicted score is interesting once. After that, what matters is what you do about it. We tied each prediction view to a next step: a recommended practice set, a topic to review, or in some cases an explicit "you're on track, take a break." The page wasn't a scoreboard. It was a starting point for the next session.

Scheduled sessions

The AI scheduled practice based on the student's target test date and current performance. Our job was to make the schedule feel like a guide, not a prison.

We built the session card around three pieces of information: what kind of practice it was (timed section, drill set, review), how long it should take, and why it had been recommended. Reason matters when AI is doing the choosing. "Practice logical reasoning, 25 minutes" feels arbitrary. "Practice logical reasoning, 25 minutes, this is your weakest section and you haven't touched it in five days" lands differently.

Skipping a session didn't break the schedule. It rebalanced. We didn't want students to feel punished for missing a day, because most of them would miss days.

Customisable study preferences

The default schedule wouldn't work for every student. We gave them controls: study days, daily time budget, weekly intensity, preferred session types. Personalisation that the user could actually personalise.

The design tension was keeping the surface light. Test-prep apps tend to drown users in settings. We pushed almost all of it into a single preferences screen, layered by frequency of change. Study days at the top (changes weekly), preferred session length below (changes monthly), notifications and reminders below that (changes once and then never again).

Bookmarks and review

Questions a student got wrong were the highest-value content in the product. The bookmark system captured them automatically and gave students a review surface that showed the question, their previous answer, and where the reasoning broke. Review wasn't optional. It was where the learning actually happened.

The review loop was where the AI quietly earned its keep. The model was prioritising which wrong answers to surface next, weighting by how often a similar mistake had been made, how long it had been since the student last touched that topic, and whether the underlying concept showed up frequently on the test. The student didn't need to know any of that. They just needed the next review queue to feel sensibly chosen.

Brand

Branding sat alongside the product work. The direction we landed on blended law with a softened futurism. Serif typography gestured at legal tradition. Geometric forms pushed against the textbook feel. The colour palette was saturated enough to read as a tutor's voice rather than a study guide's. Not stiff, not gimmicky. Closer to a coach than a workbook.

What I took from this project

Shaipe shipped its core features into student testing. The work translated AI diagnostics into a study rhythm a user could follow without needing to understand the model behind it.

The pattern I took from this project, and have used since, is that AI personalisation is only useful when the user can adjust it. A schedule that's good and adjustable will be used. A schedule that's perfect and locked will be abandoned. The product wasn't trying to be smarter than the student. It was trying to give the student a system they could trust enough to keep showing up to.