Inspiration
Companies invest a lot of time/money onboarding new employees, especially for short term roles like interns or transitioning non-technical personnel to technical-adjacent roles (project manager). For a 12 week internship, about 3 weeks are spent onboarding, during which the company is not getting meaningful work, and the employee is not getting the experience they want.
We’re presenting an app that will cut the time needed for learning in half. By tailoring the subject matter directly to the learner, adapting to their learning style, needs, and interests, we anticipate that not only will our users learn faster, but they will be more motivated and less discouraged.
What it does
Our project takes existing research papers and detailed documents and uses GPT3 to generate two progressively higher level versions of the document. We then generate a quiz directly from the document to test the knowledge of the user, and, if necessary (when the user scores poorly), open a chat dialogue with GPT primed to teach the user more about their most-missed topics.
How we built it
We used the OpenAI API to intelligently process technical documents and extract key concepts in simplified forms to present information to the user at the level they feel most comfortable. Also using the OpenAI API, we then generated a quiz tailored to the source document that allows the user to test their knowledge using short answer questions. If the user doesn't perform as well as is necessary, a chat dialogue opens with a GPT session (also OpenAI API) primed to teach them. User data is stored and aggregated in a database to allow for integrated analysis and personalization.
Challenges we ran into
Ensuring summary document fidelity to the original document was one issue that we ran in to, but by pushing the model to focus on concepts rather than technical data, we were largely able to mitigate this issue.
Accomplishments that we're proud of
Accessing and utilizing the OpenAI API and engineering our prompts to get specific, usable outputs. Additionally, parsing the large amount of data we could collect to determine what we should collect (what was useful).
What we learned
The importance of front-end development and tweaking prompts to ensure a consistent, usable output from LLMs. Additionally, we learned the importance of building with an end user in mind and defining a vision to guide our development.
What's next for AccelerEd
In the future we want to distinguish ourselves from competitors with more AI-enabled personalization, potentially implementing ideas like:
- Aggregating and analyzing user data to optimize the presentation of information for all users, as well as tailor the learning process to previously underserved learners like those with attention disorders
- Expanding the platform for use in schools and non-professional environments, branching out from analyzing technical papers
- Creating a database of learning resources, and recommending specific resources to users based on their data
- Integrating difference forms of media (image and audio generation, LaTeX comprehension, etc)
- Integrating sentiment analysis into the chat to encourage the user and keep them motivated throughout the learning process
Log in or sign up for Devpost to join the conversation.