Reimagining Education

Inspiration

Being students, learning efficiently becomes a vital point in how we change our lives, absorbing as much information as we can in this limited time we get. Further away from us lie millions of students who are barely able to reach this content, and when they do so, end up losing more than 50% productivity due to the traditional 'one-size fits it all' model, creating high repetition and inefficient learning due to lack of compatibility. This is where the idea of adaptive learning technologies comes in and at the right year with the power of Large Language Models capable of storing knowledge beyond imagination.

What it does

In simple terms, it teaches you concepts the way you best learn from them. The way we architect it begins with an assessment to identify the topics that the person is weak at. The system simply asks questions from subtopics decomposed from the topics from which we compare open-ended answers with an ensemble of generated questions and answers. It then begins to curate classes, videos, documents such as class notes, and practice questions, with a focus on courses they find difficult to understand. This happens iteratively, testing the student and curating more content personalized to the user, as he develops his skills. Past experience with collaborative filtering allows the system to understand users' needs and respond based on them.

How we built it

Starting with ML, since it was the specialization of the members, we built the Backend pipeline, capped at In-Context learning due to the computational cost of fine-tuning one, as well as the time feasibility. Technologies include GPT3.5, and OpenAI Embeddings combined with sentence-transformers(later discarded due to immense memory consumption, substituted by plain old numpy). Following this, these methods were exposed over Flask APIs linked through Firebase and were queried by the ReactJS front-end. As a final touch, we host the backend on Heroku and the front-end on Vercel.

Challenges we ran into

Bugs. A lot of bugs. And on the WebDev side due to the lack of our experience in it. Computation capped our capabilities to wield ML models to their full potential but managed to settle with OpenAI calls. Prompt engineering was very sensitive and required precise handling of the text, especially fixing the generation format to a fixed one to ensure automated processing.

Accomplishments that we're proud of

The entire ML pipeline works at a perfect pace with close to 100% accuracy for its generation capacity. The backend and front end adopt BLoC architecture and connect successfully through a hosted server.

What we learned

How to hack up a pipeline in a day's time, putting together frontend, backend, database, and ML into a singular working project. Learned new tech stacks, both frontend and backend, along with prompt engineering. Collaborating over Github on a high-frequency commit rate, taking on different tasks, and finally bundling them all together.

What's Next for Redefining Education

We build the product with endless possibilities in mind.

• The AI can be improved by fine-tuning, Retrieval-Augmented Systems, and several other tricks
• The assessment will possibly be diverse
• Creating several tracks to learn, with several topics within each track have been embedded into our code, yet to be used
• The number of learning methods is huge and sadly, only a few could be implemented through the weekend. These would improve user performance and facilitate the product's vision.
• Utilization of Collaborative filtering at a large scale to analyze learning methods, the pace of learning, etc.

Built With

• firebase
• flask
• generative-ai
• heroku
• huggingface
• large-language-models
• python
• react
• vector-embeddings