NOTE FORGOTTEN IN VIDEO: RabbitHelper solves across a range of disciplines, not just arithmetic problems from a deterministic set. It can dynamically adjust solutions such as the .jpgs linked in the extra additional info page.
Inspiration
Proactive personal agents are everywhere now, but none of them help with personalized, long-term studying. That's one of the most valuable places proactive AI could go, and also one of the hardest, because everyone learns differently. So we built RabbitHelper to be a fun, spontaneous sidekick who keeps you on track and makes you smile while you work. After all, who doesn't want an extra study partner?
What it does
RabbitHelper is a study companion built for iPhone Duo, designed around the way students actually do math: by hand.
Opened flat, the Duo becomes a sheet of paper. You write your work with your finger, or import a photo of your notebook, and RabbitHelper reads along. When you're stuck, you fold the phone like a book. RabbitHelper dives into his burrow and pops out on the facing page, right beside the line where your work went wrong, and circles the exact spot in chalk.
He tells you where the mistake is, never what the fix is, so you still do the learning. Fix the line and the circle disappears; finish the problem and he celebrates. Works offline or online.
Why it needs iPhone Duo
On a regular phone, feedback has to cover your work or replace it, so you can never see your working and your tutor side by side at a readable size. The Duo gives your work a full page and your tutor the facing one, and the fold itself is how you ask for help: no button, no switching apps. It's the same gesture as holding your notebook open for a friend to look at.
How we built it
Built for the fold. We use iOS 27.1's new iPhone Duo APIs. The work page is anchored to the fold using reserved regions, even while the phone is flat, so folding never moves or rescales your writing, and nothing ever spans the crease. The hinge API triggers RabbitHelper's dive between pages, following Apple's guidance to use the hinge for interactions and reserved regions for layout. Tabletop and closed poses are designed too, not just the hero pose.
Reads handwriting on-device. Your ink or photo goes through Apple's Vision text recognition, which gives us every line of your work and the position of each number on the page.
Finds the mistake without giving away the answer. Every valid algebra step still has to hold at the true solution, so the first line that doesn't is the mistake. If you write \(3x + 5 = 20\) and then \(3x = 25\), RabbitHelper circles the 25 and asks what happens to the 5 when it crosses the equals sign. The judge reports only which line it is, where on the page, and what kind of slip (sign or arithmetic), so the fix can never leak. Nudges climb a ladder from a gentle question to more direct help, and they never state the answer.
A world, not a widget. RabbitHelper is a hand-drawn pixel-art sprite with dozens of animated states. He lives in a little Wonderland scene on the facing page, dives through his burrow when you fold, and circles mistakes in chalk.
The brain behind him. Photos, PDFs, and handwriting become structured problems with private answers. Hints come from a multi-tier model chain (cloud LLM → on-device AI → scripted fallback), and every hint is leak-checked before RabbitHelper says it. He also remembers each learner across sessions.
Individual contributions
Ben owned everything the app knows and says. He built the pipeline that turns imported photos, PDFs, and handwriting into structured problems with private answers; the struggle tracking; the leak-checked hints from a multi-tier model chain (cloud LLM → on-device AI → scripted fallback); and cross-session learner memory. He validated the system with headless test runs and simulated hardware inputs, exposing state cues for the interface through a single boundary without touching any visual or layout code.
Jonathan owned where and how RabbitHelper appears on screen. He built the interface across every device pose (flat, tabletop, book, and closed), using reserved regions and hinge signals to drive smooth layout transitions, sprite animations, and character placement. He built all the interactive views and controls, and verified pose changes and gestures with stubbed data and UI test suites, independent of the live coaching logic. Owned on-device, offline fallback brain logic.
Challenges we ran into
Most of our difficulties came from designing long-horizon workflows without real student work to learn from. Our harness suffered from context loss, memory leaks, and poor instruction following. We resolved most of these because we come from very different technical backgrounds, and talking problems through led to plenty of "eureka" moments.
Building for brand-new hardware was its own adventure, since the Duo's fold APIs had only just shipped and we were learning them as we built.
Accomplishments that we're proud of
- We turned the fold into a natural way to ask for help.
- We built a tutor that reads your handwriting and circles the exact mistake without ever giving away the fix.
- We gave it a pixel-art world with a character you actually want to study with.
Most of all, we built something we'd happily use every day. In fact, he oversaw this submission.
What we learned
Everyone learns at their own pace, and using AI to figure out exactly where you are is the best way to find the rabbit holes that actually interest you. We also learned that new hardware makes new interactions possible: a phone that opens like a book turns out to be a natural home for a study partner.
What's next for RabbitHelper
We want to give RabbitHelper a voice, deepen his long-term memory of how each student learns, and extend his judge beyond algebra, from chemistry to essays. Apple Pencil support will make writing on the Duo feel like real paper.
As knowledge becomes more of a commodity, the process of learning, and the mental flexibility it builds, matters more than ever. Our goal is a study buddy that protects that process and stays with you, fold after fold, until you finish your education.
Log in or sign up for Devpost to join the conversation.