AI exams, AI marking and study notes in a mobile learning app

Problem

Students on this AI learning platform needed three things. None of them worked like a normal app screen.

• Practice on their own topic. A fixed question bank cannot cover every class, subject and chapter. It also cannot use the notes a student already has.
• Marking for long answers. Creative questions are the long-answer part of school exams in Bangladesh. Each one has a short passage, then questions that move from knowledge to higher-order thinking. Students answer them by hand and wait days for marking.
• Notes at the right level. One lecture note is too detailed for some students and not detailed enough for others.

All three depend on AI that is slow and sometimes fails. If a student taps Submit and sees nothing, they will tap again or leave. And if leaving loses their work, they will not come back.

Solution

I led the mobile team of three developers on this app. For these three features, I designed the flows, the rules and the screens. I split the work across the team and reviewed every pull request. The developers wrote most of the code to that design. The backend team built the AI generation and marking APIs.

Exams on the student's own topic. A student starts from a preset, from a class, subject and topic (multiple choice, descriptive or mixed), or from an uploaded document. Every exam moves through the same states. A preparation screen shows while questions are generated, and the student confirms before the clock starts. The result ends with a learning path: topics to revise, each with the source to revise from. A retake creates a new exam instead of repeating the same questions.

CQ papers that do not lose a page. Each sub-question can be typed or answered with photos of handwritten pages. A photo uploads as soon as it is added. Each finished question set is saved to the server as a draft, so the work comes back if the student leaves.

The paper has a countdown and blocks the back button. It cannot be submitted until the required number of sets have an answer. During submit, the app shows each stage, with a Retry button if it fails. While the AI is still marking, the result page checks again every few seconds. Results split the score by knowledge, comprehension, application and higher-order questions. Each sub-question gets feedback and the book references behind it.

Notes at three depths. Each lecture note has three versions: 5E, Easy and Comprehensive. The Easy version explains the topic as if the reader is five years old (ELI5). Easy and Comprehensive are generated only on request, and only once. A second tap is ignored, and the student cannot leave the screen until generation finishes. From a note, a student can open its worksheet or its exam. An existing one is reused, and finishing the exam takes the student back to the note.

What I built myself. Later, I redesigned the exam hub around exam formats. Each format has its own page, with a confidence meter and past exams by status. I built that redesign myself. I also built the link that opens the AI tutor with one specific CQ sub-question already attached.

The trade-offs. A CQ draft is saved when a question set is finished, not on every keystroke. An unfinished set is not on the server yet. But every save is a complete answer with its photos. And the app does not send a request for every word over a slow mobile connection. Generated exams and notes also mean a wait that a question bank would not have. The app shows that wait instead of hiding it. And no AI cost is spent on note versions that no one reads.

Impact

The platform is in pilot testing, so there are no usage numbers to report yet. What changed is how the app behaves:

• A student's finished work is kept if they leave the exam or close the app.
• Only papers that follow the exam's rules are sent to AI marking.
• A slow or failed AI step is visible and recoverable, never a blank screen.
• Every result ends with what to revise and where. A note leads directly to practice and back.