PeachDeck

What it does

PeachDeck turns PowerPoint slides into AI-generated video presentations. It doesn't just summarize the slides, but it uses RAG LLM model to expand upon the lecture and teach the nitty gritty of that professors didn't mention in only a matter of minutes. It can be applied for consultants, training videos, teachers, and students.

Inspiration

What happens when you miss a class and all you have is a deck of lecture slides filled with bullet point content and no context? One of our members often missed classes because of illness/other commitments, and they fell behind because lecture slides often provide insufficient context. What if there was a way for them to catch up on lectures easily? Using RAG LLM, however, we can turn disjointed lecture slides into a cohesive video presentation with immaculate generative voiceovers, allowing students to pick up missed concepts in a short amount of time. But our AI-powered journey doesn't stop there. We're not just filling gaps; we're amplifying the learning experience. Instructors often can't dive deep into every concept in class, leaving students curious about what wasn't covered. Our AI agent steps in, providing in-depth explanations for concepts left untouched in class. It's not just about catching up; it's about delving into the intricacies of each subject. The use case is not confined to schools alone – it can help with any kind of training and educational videos. It reduces the time it takes for instructors to prepare online courses, and also eliminates the need for extensive manual video production, making corporate training more budget-friendly. We hope to democratize access to high-quality content, reaching learners around the world and contributing to a more equitable education landscape.

How we built it

Use Bun to create the React app and install all needed packages Used PredictionGuard AI from Intel to create a safe RAG LLM model (AI Agent) for university textbooks Used Python for data cleaning We found open-source textbooks with free for commercial use licenses which we then embedded as vectors with multimodal CLIP We used the open-source en_core_web_md model from spaCy to classify what subject a lecture belonged to and thus indexed to the appropriate vector embedding space Used ElevenLabs for audio generation with multiple generative AI voices Used React for the frontend Used Figma for prototyping: https://www.figma.com/file/TPEmQIIjMAyELI5sIkzbeP/Trees!?type=design&node-id=93%3A26453&mode=design&t=ruoBvc9hY4f9oCzR-1 Used Flask for the backend server Used Spire to convert pptx slides to images and add audio over them Challenges we ran into We ran into many challenges when merging our front-end and back-end programs together, as the merge involved combining a variety of difficult conversions, such as extracting text from powerpoints, integrating text with audio, and converting audio to video. The last one took an especially long time to resolve as generating video files are often quite tedious. Asking prompts - after 4 hours of playing around, we discovered that LLMs did better with less information than with more counterintuitively

Accomplishments that we're proud of

We are especially proud of the level of the pipeline we've managed to execute, using a plethora of different algorithms to make something extraordinarily useful for the education industry. End-to-end design of the web app on Figma after many iterations; we want to create something that is aesthetically pleasing and intuitive to use I have never built a RAG LLM model before and was surprised to learn how cool and fun it was, especially with mentors from Intel!

What we learned

Getting advice from experts in the field of what you're looking into is extremely helpful. By talking to experienced engineers and thinkers early on in our journey, we were able to dream big and accomplish tasks at a more monumental level.

What's next for

We hope that the product would evolve into something that allows users to edit the AI-generated content and customize it more directly. Users can review generated scripts from slides first, and regenerate if they are not satisfied. They can also edit the video directly within the platform, and use the chatbot for further assistance. Ability to add AI-generated avatars (we already have voices and the ability to change between them but do not have a clear way to do this on our current interface).

In the future, we also want to add translations, so it enables people who are not fluent in English learn from slides in their native language! We hope that one day everyone can attend any lecture in the world!

Built With

agent ai api bun css elevenlabs figma flask html intel javascript langchain llm neurachat numpy open-source pandas python rag react spacey spire