Learnery

Inspiration

Nearly 300 million students in the US alone have some sort of learning disability. No platform exists online to deliver lessons to such students with multiple modes of delivery that delivers it in the way best suited to each student.

What it does

Our platform enables students to learn curriculums in multiple different ways, and tracks their progress in the different modes of deliveries and the different subjects that they are studying. Based on this data, it then recommends the next lessons to deliver to the student, and also the best mode to deliver them in.

How we built it

We built Learnery by building a LSTM neural-network using PyTorch, which interact with our FAST API Python Backend, connected to our frontend built in TypeScript, React, and Next.js. More details on how the neural network works and was implemented can be found below.

Overview

The backend was built with FastAPI and Python. It provides Learnery with an API for feeding in inputs, such as student performance data and profile, and accessing the resulting predictions of the recommendation engine for what lessons should be done next.

The following neural network architecture diagram outlines what inputs would be used and what recommendations would be produced.

neural_net

We have implemented a neural network in Python with PyTorch and wrote scripts to generate labelled mock data to train on. In essence, the recommendation engine will help the user tune up difficulties in types of lessons they are doing well and have a record of doing well in, and moderate the frequency of lessons across different categories. This also encourages the user to grow in their weakest learning dimensions. For example, a child with historical performance data suggesting they’re struggling with numeracy and literacy would have those lessons be recommended at a higher frequency with a suitable level of challenge.

How would we train the model in practice? (Post-hack viability) To overcome the ‘cold start’ problem, we can implement a non-AI solution for the recommendation engine. For example, one powerful approach is to use a collaborative filtering algorithm in the early launch stages of Learnery. After sufficient data has been collected, we can consult experts and produce suitable labels to train the engine to make predictions as experienced therapists would.

Post-hack viability

As mentioned above, we have planned how we would generate the data needed to get off the ground. In addition, Learnery also supplies a user-friendly CMS using Sanity.io that we implemented. Through this, in the future therapists will be able to contribute reusable lessons and share them with all users of the platform. therapist_page

Challenges we ran into

Implementing a complicated neural network while also a polished frontend was a massive technical challenge to achieve in this time frame, but an extremely rewarding one.