Inspiration
Personally, I often find English reading comprehension challenging especially since it is not my first language. However, I find traditional comprehension practice questions making me unmotivated, so I decided to create a programme that can generate comprehension practice based on topics that I am interested in.
What it does
The LLM generates a short informational passage about a random topic chosen from a list (which you can add to through "add your own topic" input), then generates a question about the passage. Upon user's answer input, the LLM evaluates user's answer by comparing it to an "model answer" generate by itself and returns either "correct" or "incorrect" based on its judgement.
How we built it
I started with researching about pure front-end AI models that runs in javascript, and decided on Transformers.js. Then I did some prototyping with different models available in javascript to get to know how to use the main functions. After that, I created minimal html and css for simple inputs and outputs, together with the fully functioning javascript for the whole text-question-answer process. Finally, I use Gemini to style the html.
Challenges we ran into
AI models are usually run on back-end servers, so it took a long time to find a working source for Transformers.js that does not require a server at all. Also, since the LLM used is really compact for efficiency, it was really difficult to prompt it to make it consistently produce desired responses.
Accomplishments that we're proud of
First time implementing AI into a programme.
What we learned
The super large variety of specialist AI models other than mainstream LLMs, like object detection in image, image-to-image, etc.
What's next for AI Comprehension Trainer
Add in more customisable features, such as reading grade level, question type, difficulty, etc. Add in different modes like practice / exam mode.
Built With
- css
- html
- javascript
- qwen2.5
- transformers.js