Storyverse

Inspiration

Sankar came up with this idea while he was trying to figure out how to inspire a love of reading in his child. Unfortunately current educational systems are one-size-fits-all, and he was running into educational materials with themes that didn’t match what his child found exciting.

What it does

Storyverse uses Generative AI to enable educators to create personalized content for each student that both engages at a high level, and meets the targets of each exercise. There are 3 parts to the application,

1. First we let the Educator input parameters for the exercise. Today, we let them input the age of the Student, and the theme that they want to use, for example ‘Baseball Championship’ or ‘Space Invaders’.
2. Next, we used Generative AI to create a multi-page story for the Student, where each page has an illustration for the scene that was also built using Generative AI
3. Finally there’s a quiz page, where users are presented with a question, respond in plaintext, and get feedback from the LLM.

How we built it

To build this, we used the following technologies,

• LangChain was used to establish a prompt-engineering pipeline that resulted in a prompt for generating the stories and quizzes that can take in variables like age and theme.
• Anthropic’s Claude is the LLM that used the prompt we created to generate both the story content and the quiz. It was also used to determine whether student answers are correct or not, and to provide feedback as to why.
• OpenAI’s DALL-E 2 was used to generate images based on each scene and quiz response to further engage students.
• Our frontend was built with React / Typescript, with a Python / Fast API gateway, hosted on Vercel

Challenges we ran into

Unfortunately the time required to fine-tune a model made it impractical to do within the 8hr window of the hackathon.

Accomplishments that we're proud of

It works - after a 1-day hackathon no less! As many of us are parents that are actively trying to teach our kids outside of the classroom, this seems like a powerful flow that we can curate for our own needs.

What we learned

There is a dance to navigate regarding the tradeoffs between fine-tuning vs. prompt-engineering and additional context, and when to bundle into a single prompt vs. when to chain queries. We chose the hackathon path, but we’d love to do more testing.

What's next for Storyverse

Battle testing with our kids for product discovery, and maybe even shopping around educators to see if there’s real business potential here.

Built With

• anthropic\claude
• fastapi
• jupyter
• langchain
• openai\dall-e-2
• python
• react
• typescript
• vercel