Tutor Assist

Inspiration

Our idea is to leverage AI tools like chatbots to personalize education for students. Every student has a different learning speed and ability to grasp complex topics. With AI tutors, lessons can be tailored to match each student's needs. The AI tutor can ask questions and assess the student's current knowledge and then provide explanations and examples suited to that individual. As the student works through material, the AI can track their progress and adapt the lessons accordingly. If a concept is too advanced, the tutor can break it down into simpler steps. If the student is breezing through, it can provide challenges to go deeper. This type of 1-on-1 personalized instruction allows students to learn at their own pace. AI tutors have the potential to revolutionize education by ensuring no student gets left behind. The possibilities are exciting as these technologies continue to develop. This project explores Claude's abilities in math tutoring and natural language processing.

What it does

We have developed a Tutor Assist app. It starts by the tutor specifying the topic and details of personalization required. It leverages advanced natural language generation to produce personalized training plans, introductory content, supplementary materials, and comprehension questions - all tailored to the student's needs.

The materials are designed to be engaging and educational, tapping into Claude's vast knowledge base across subjects. Tutor Assist continually refines its approach through real-world use, ensuring the materials are optimized for learning outcomes.

With Tutor Assist, tutors can spend less time on busywork and more time connecting with students. It reduces prep time while providing each student with a customized learning experience. The app ultimately allows us to leverage Claude's capabilities in language understanding and conversational education.

How we built it

After signing up for Partyrock, my team and I spent a few hours reviewing the platform to understand its capabilities and limitations. We looked at some existing apps to see how they were built. We also experimented with the App Generator and read through the Partyrock guide and prompt engineering guidelines.

One interesting capability we noticed is that Amazon Bedrock seems able to handle mathematical reasoning to some degree. We were curious to explore how well it could limit hallucinations, so we decided to build a basic math tutor app as a test case (https://partyrock.aws/u/aicode/EzWR650p-3/MathWhiz).

We started by drafting the conversational interface and dialogue flow for the math tutor app. This allowed us to design the overall user experience before diving into the AI capabilities. Overall, our initial exploration gave us a feel for Partyrock's potential, though we still need to build out the full app to truly evaluate its strengths and weaknesses.

Then we picked up a tutor assistant use case which has the potential to leverage more of Partyrock's capabilities and provide production ready use case for our tutors.

Challenges we ran into

When building our math tutor app, we started with an empty Partyrock app and gradually added widgets based on our conversational interface design. Originally, we had hoped to incorporate generated images to enhance the user experience. However, we found that image generation was too slow for our real-time use case, and actually detracted from rather than added to the experience. Given this limitation, we decided to focus on text-based conversational interactions without images for our particular math tutoring app. The slow image generation speeds may work for other use cases, but for our needs, it led to a poor user experience overall. By keeping the app text-based, we were able to provide a smoother math tutoring experience.

We encountered some cases where the Generative AI hallucinated and did not produce the correct result, as can be expected with this type of technology. However, through refinement of the prompts, we were able to significantly reduce these types of errors.

By the time we were ready to build our Tutor Assist app, we have learned enough and the 'Generate App' worked perfectly well. Our actual system time was less than 2 hours to make it fully ready and useful!

Accomplishments that we're proud of

We were pleasantly surprised by how easy it was to create an application using generative AI. The ability to build and iterate entirely in the cloud and browser meant our team could collaborate rapidly without any local setup.

What we learned

Another notable capability we observed was the model's ability to provide meaningful intermediate feedback to students, like suggesting areas for improvement and providing the right answers, even when not explicitly prompted to do so. This flexibility highlights the power of generative AI to understand context and engage in thoughtful dialogue.

Overall, the simplicity of deploying generative AI services and the robustness of conversational abilities exceeded our initial expectations. These strengths should continue enabling faster and more capable application development going forward.

What's next for Tutor Assist

Tutor Assist is our starting point in creating different teacher personas. A key to our AI tutors to connect with students is understanding the diverse learning styles, interests, and backgrounds of the student population. These personas should be equipped with relevant materials and teaching approaches that can be tailored to meet curriculum goals and connect with each student's needs. Generative AI can help automate the process of analyzing student data patterns and producing personalized teacher profiles. But human oversight is critical to refine the personas and ensure appropriate, high-quality learning materials are provided. The end goal is to give each virtual teacher persona the content and tools to optimize student engagement and outcomes, while aligning with educational standards.

We plan to develop a full set of educational content on a specific topic to demonstrate how the same topic can be delivered in different ways based on student needs in the next quarter

Built With

• amazon
• claude
• partyrock