Tutera.ai

Inspiration

The inspiration that drove this project comes from the personal experiences of everyone on the team who has gone through classes where access to a tutor was limited or unavailable. We have also experienced the state of education in public schools where classes are filled with 40+ students and only one teacher. The lack of one-on-one instruction is costing students the ability to achieve their full potential.

What it does

Tutera.ai is designed to be a personal tutoring chatbot that learns from material provided by instructors and guides students step by step through problems, developing their critical thinking and improving their academic achievement.

How we built it

Using the AutoGen multi-agent framework, we were able to use prompt engineering to achieve agents that maintain consistent roles in tutoring and instructing students. We used next.js for creating the user interface and flask for hosting the application locally.

Challenges we ran into

One major challenge we ran into was that we could not use the intel developer cloud resources to train our object detection to OCR model. A key component in our data pipeline, in order to ingest textbook PDFs, we needed to train the object detection to OCR model on our annotated dataset. We spent hours trying every possible fix and even spent time with the Intel team trying to go through fixes to no avail. It turns out the software we were using was not compatible with the version of Pytorch and the Intel environment.

Accomplishments that we're proud of

We are proud of working with AutoGen to create a multi-agent framework that did not require any fine-tuning. This was useful as tutoring dialogue is very scarce and often is structured differently based on the subject.

What we learned

We learned a lot about multi-agents and how they function. Attending workshops by companies such as fetch.ai really helped us achieve our goals.

What's next for Tutera.ai

The next step for Tutera is expanding the multi-agent framework to add agents that can use tools to improve the model's output accuracy and reduce false information. We also want to add a rag agent that can retrieve the textbook's PDFs and reference specific information in the textbook, passing that information to the tutoring agent.