Engineering · Applied AI

Where the LLM sits in a question bank: seven jobs, and the one it is not allowed to do

A Malaysian tuition provider's content platform uses language and image models to read textbooks, write questions and learning cards, tag them, review them and redraw figures. It is never allowed to publish. A tour of the pipeline.

Pikabyte Engineering 4 min read

9,300+

questions produced through the pipeline, each reviewed before it can be published

“We use AI to generate questions” describes almost nothing. In the content platform we built for Tavis, an online tuition provider for SPM students in Malaysia, a model is involved at seven distinct points, each with a different job, a different risk and a different check on it. At the time of writing the platform holds more than 9,300 questions and 5,600 learning cards in English and Bahasa Melayu, drawn from 475 uploaded textbooks and papers.

Here is what the model does at each point, and what stops it going wrong.

1. Reading the textbook

Subject teams upload textbooks and past papers as PDFs or images. A model reads each page and returns two things: the text as clean, structured content, and the location of every figure on the page.

The second part is where a detail matters. The model does not produce the figure. It reports where the figure is, and ordinary code crops that region out of the original page. The diagram a student eventually sees is the publisher’s diagram, not a model’s recollection of one.

2. Sorting pages into chapters

Extracted content is mapped to the syllabus: this section belongs to Chapter 5, that one to Chapter 6. The result is a knowledge base organised the way teachers think about the subject, which every later step draws from.

3. Writing questions

A generation job asks for questions on a chosen subject, level and chapter. The model is given that chapter’s material from the knowledge base and writes from it.

This grounding is the main defence against the classic failure of AI-written content: questions that are fluent, plausible and not in the syllabus. The model is not asked what it knows about the water cycle. It is asked to write from these pages.

Each question type has its own strict structure, such as options and an answer key for multiple choice, and anything that does not fit is rejected by validation. Every question is produced in each language the subject requires.

4. Tagging by subtopic

Questions are tagged with the subtopics they test, chosen from the chapter’s own list. An editor can also ask for suggested tags on a single question. Suggestions are exactly that: nothing is saved until a person accepts.

5. Reviewing its own work

Before an editor sees a generated question, a review job casts one or more independent AI votes on it. The number of votes is configurable. Each vote judges the question good or not, and may propose a corrected version.

The rule is strict. More than half the votes must be good for the question to be approved automatically. A tie fails. Anything that does not pass is marked for human review.

This stage is not a rubber stamp. On the live dashboard, thousands of questions and learning cards sit in a queue labelled as flagged by an AI reviewer and waiting for a person. That is the system working as designed: the model sorts the pile so that editors spend their time on the items that need judgment.

6. Writing learning cards, then drawing, then checking

Learning cards pair a term with an explanation and an illustration. They are made in three separate steps.

First a text model writes the card and chooses which of the chapter’s textbook figures best illustrates it. Then an image model colours or redraws that one figure in the house style. Then a separate check compares the result with the source figure and reports where they differ.

Splitting it this way keeps each step checkable. A single “make me a card with a picture” request would give a result nobody could audit.

7. Redrawing figures on request

When a figure needs a cleaner version, an editor starts a regeneration job. The model produces a candidate. The editor sees it beside the original and chooses to apply or discard it. The original is preserved either way.

The job it does not have

The model cannot publish.

An automatic approval marks a question as approved inside the platform. Sending it to the client’s live learning product is a separate, explicit action taken by a person, one item or one batch at a time. Images are copied across with it. Until someone presses that button, students see nothing.

Everything else in the design follows from that line. AI can draft, sort, suggest and check. A named person decides what a student reads.

What makes this hold up in daily use

  • Prompts belong to the content team. Each stage has its own instructions and model setting, grouped into sets that can be assigned per subject and edited in the application. When the science editors want different colouring rules for diagrams, they change them without a developer.
  • Long jobs report progress. Generation and review run in the background and stream their status to the screen. A failed review can be retried and picks up where it stopped, keeping the votes already cast.
  • Corrections are shown as differences. A proposed fix appears word by word against the original, so accepting it takes a moment.
  • The contract is tested. More than 1,150 automated tests cover the backend. Models change. The rules around them should not.

What a business outside education can take from this

Replace “question” with “product description”, “contract clause” or “customer reply” and the structure is the same.

  1. Break the work into steps and give the model one job per step.
  2. Ground each step in your own material.
  3. Let code handle anything that must be exact, such as cropping an image or validating a structure.
  4. Have the model check its output, and send disagreements to a person.
  5. Keep publishing as a human action.

An AI feature built this way is slower to demo and much easier to trust.

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

The work behind this article

Platform · Education · Assessment content

Textbooks in, reviewed exam questions and learning cards out, in English and Bahasa Melayu.

A content production platform for Tavis, an online tuition provider for SPM students in Malaysia. Subject teams upload textbooks and past papers, AI extracts and generates questions and learning cards against the syllabus, an AI review stage votes on each one, and editors approve what gets published to the live learning platform.

9,300+
questions generated and tracked
5,600+
learning cards, in English and Bahasa Melayu
475
textbooks and papers read into the knowledge base
1,150+
automated backend tests
Tour Question Bank AI Tavis · Malaysia

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.