YouQuiz GPT

I got on local news: "SuperHaber" with my project and presented it.
Logo of YouQuiz GPT

YouQuiz GPT

First off before starting to explain my project, I want to send my deep gratitudes to the organization team behind this hackathon for providing me a chance to show my skills and present my project as a highschooler AI enthuasist. I especially believe this project best fits the AI indiviual learning category and generates real world impact.

My inspiration to create this app came to my mind in school during the breaktime. In school we don’t have access to internet so I wanted to creat an assistant that uses local LLMs to generate quizes with RAG system. I wanted to watch the videos that I saved at home and than take quizez on specific parts of that lecture video. That is how I came up with the idea of YouQuiz GPT and now my whole class uses it. Even my teachers while giving out assesments.

Furthermore I want to briefly explain the workflow, you also may investigate the code that I provided in the github repository. First it transcribes the voice of a youtube video into a text with Whisper model. Then that text is seperated into chunks and then embedded so that gpt-oss can process the text. Further, the vector database and qa chain get locally generated. Afterwards with a complex prompt which also relies on inputs from the user, an output in JSON format is generated. Finally the raw LLM output gets turned into a fully functional quiz UI where users can even check their scores in the end.

Now I want to point out the usage of gpt-oss:20b in this paragraph. I used Retrieval Augmented Generation (RAG) AI technique which is improving day by day. As you can see from my other repositories on Github, I am an enthuasiast about artificial intelligence and love implementing it to daily life. I also have experience in data analytics such as gathering necessary metrics and analyzing data. With this knowledge I tested different optimizations (temperature, top_k, top_p, context windows, etc.) and determined that the best output is via the current settings of gpt-oss:20b. I have noticed during my QA tests that gpt-oss has an incredibly strong grasp of long-range dependencies which was key in my project. Also when compared to other open source models gpt-oss consistently maintained coherence across varying contexts and always provided well-structured logical quizez. In over 200 hundred tests that I have noted, not even once did gpt-oss:20b give an output that wasn’t in JSON format. I have also added my initial version before this hackathon to Github, in that version as you might check, I have used other models and the outputs were going out of context and into hallucination. I have noticed that Ollama’s nomic-embed-text embedding combined with gpt-oss:20b is unstoppable in case of RAG systems as I further explained in my video showcase. So that is how gpt-oss:20b is an open model that has fully shaped YouQuiz GPT as well as Ollama.

During the coding phase I ran into a few challanges. My main challenge was writing the base prompt for this application. I signed in for a prompt-engineering course by Vanderblit University and used Question refinement patterns to achieve the best prompt. I also decided to implement the persona pattern and chain-of-thought prompting into my prompt as well to get the best results. However in my opinion the best strategy I used in prompt engineering was the input based system I generated. I made it so that the question-number/ difficulty/ language/ context user inputs should all be added into the prompt since gpt-oss:20b is a master at outputing on context. I even gave explanations to different difficulties to further help the reasoning of gpt-oss:20b. You can view the full prompt and the RAG system in the main_minimal.py code in my repository.

If I manage to learn more about programming UI and improving UX, I believe the future of YouQuiz GPT is really bright. Also I need to get a new computer to test LLM models and langchain modules with higher RAM requirements such as the gpt-oss:120b model which I unfortunately never got the chance to try. Furthermore I am planning to implent vision-based AI systems to analyze youtube videos with their visuals as well. If I work hard and overcome these obstacles I want to launch YouQuiz GPT as an offline mobile and desktop application! There are no limits to what YouQuiz GPT can achieve in near future. A program used by schools and courses to create assesments, a modern fun way to teach babies, training for jobs getting more interactive and so on. You imagine it YouQuiz GPT generates it. Waiting for your judgement!