SAT GPT

Inspiration

The inaccessibility of SAT prep for many low-income families led us to create this product. Also, popular applications like GPT-3 and ChatGPT cannot accurately answer SAT problems as they are specialized and struggle at MCQ's in general.

What it does

Takes in user input of questions and question type before using the updated GPT-3 to get an accurate response to the question.

How we built it

We used prompt engineering techniques and lots of experimentation to find the right template prompt to query the model. One of our greatest breakthroughs was asking the model for synonyms and trying to match them up with the answer choices. This enabled the model to use its previous knowledge to better answer Multiple Choice questions.

Challenges we ran into

Accuracy and getting the AI to work with the program was a huge stump that we still faced even during our demo. The model can be volatile sometimes and not give the same response every time.

Accomplishments that we're proud of

We think that our new approach can unlock the true question-answering capabilities of Large Language Models and enable them to give even specialized answers. We're proud of incorporating another API into our product to better the user experience. We also worked well as a team under pressure to deliver the submission.

What we learned

Website bases and how to incorporate other aspects of design into them. We used NextJS and OpenAI's API and ConvertAPI (https://github.com/ConvertAPI/convertapi-js) to enhance our product, which is a great stepping-stone to enhancing our knowledge of the vastness in this field.

What's next for SAT GPT

Adding onto the PDF reading process by creating a parsing mechanism in which the user could get the specific passage and question they need can simplify the app process. Also expanding the types of questions we fully support can allow students to get a more holistic learning platform. Finally, providing a reliability score that tells the user how much they should trust the answers and explanations of the model can increase the quality of our product.

Built With

• gpt-3
• nextjs
• openai
• react