Inspiration
Working as a TA for a number of graduate and undergraduate classes, its clear that the most common request from students is workable examples and feedback to prepare themselves for exams. These are time consuming to generate for teaching staff, so are often ignored. Students also often come to courses with different proficiencies in the topic, making some practice problem sets too complex to be helpful. We need a solution which can engage each student where they are, without occupying too much time from instructors.
What it does
Gemini Tutor allows for students to select courses in which they're enrolled as well as their study goals for these courses. We then collect background data regarding this course, and use this context and the implicit knowledge of Gemini to generate questions to gauge the student's proficiency. Then an iterative series of questions is generated to work the student towards their goal, providing study recommendations along the way.
How we built it
Gemini Tutor is built with Python via Flask. A lightweight web interface provides structure for data to be fed into Gemini, and processes its outputs via the Python backend. User data is stored via Firebase.
Challenges we ran into
In the short time, we didn't choose to build in asynchronous processing, which has caused some challenges in serving data while the model processes inputs.
Accomplishments that we're proud of
We believe functionality of the final product is impressive in its own right, and demonstrates a fair amount of promise. This system doesn't take away from the student-instructor relationship, but rather provides students with greater agency in their ability to teach themselves, which is a key component of higher education.
What we learned
We've discovered that Gemini can be a tool for learning, not just productivity. The impressive knowledge of these models can synthesize a user's current expertise level, goals, and its general knowledge about a topic in ways beyond our expectations.
What's next for Gemini Tutor
We plan to continue developing Gemini Tutor. One of the immediate goals is to implement asynchronous processing to allow for a more streamlined user experience. Additionally, we hope to explore using Gemini 1.5's long context window to ingest full lecture audio or transcripts, enabling us to provide questions better tailored to an individual course, along with comprehension questions tailored towards each lecture.
Log in or sign up for Devpost to join the conversation.