Our project is built using Bidirectional Encoder Representations from Transformers (BERT), so it made sense to follow in the tradition of Muppetware and name the project accordingly. In 1971 Grover sang a song where he worked hard to answer questions. Grover didn't quite get the answers correct, and we think Gleep is a better fit for this role. Gleep was introduced in 1967 as a Grover prototype and is a little more unkempt than Grover would later behave. Gleep knows the frog does not have to sing alone and is always ready to help.
After Columbia switched to remote lectures in March, our screen time has increased more than we would like to admit. Even by data scientist standards, it's gotten a little concerning.
We needed a better way to find information from lectures. Instead of searching keywords in slides and guessing timestamps, we developed Gleep.
Paste or upload your text, and let Gleep go to work. Get quality answers from Gleep's Q&A model or try summarization on lecture transcripts, textbook chapters, articles, and more.
Gleep uses a compressed version of BERT, MobileBERT, that runs 4x faster and has a 4x smaller model size to answer questions based on a given passage's content using script tags directly in browser.
Lecture recordings and transcripts aren't perfect. Unclear audio, delay, um's, ah's, and pauses all make it tough to follow. Summarization using BERT allows Gleep to sift through and find key phrases.
Gradio facilitates customizable UI components around TensorFlow models and Python functions. Gradio enabled our text summarization model, which we were unable to implement with tfjs scripts.
Cloudfront and S3 enable Gleep to quickly serve requests and update our models without downtime.