Inspiration
Course selection at UW-Madison has always been a problem. From unclear requisites to a complicated course selection process, the whole system needs help.
What it does
BadgerBot quickly ranks courses using fine-tune Mistral model and generates course recommendations using retrieval augmented generation.
How we built it
Analysis: Weak supervision with GPT 3.5 turbo to train a multi-layer Mistral7b model to rank courses based on preferences of different parties (students, faculty, logistics). Converted user query and ranked courses into vector embeddings using OpenAI embeddings for fast retrieval and real-time generation.
Backend: Flask
Frontend: Vanilla HTML with customized CSS
Challenges we ran into
- Collecting training data from multiple data-streams
- Model inaccuracy requiring further fine-tuning
- Bugs, bugs everywhere
Accomplishments that we're proud of
- A working product
- Successfully leveraged ML technologies to solve a real-life problem, achiving excellent performances given the time constraint.
What we learned
- Various new techniques in fine-tuning and vector embeddings
- How to collaboratively develop a state-of-art software
- The joy and importance of innovation
What's next for Badger Bot
- Integrating with course search and enroll API to achieve one-button click course planing and automatic enrollment
Log in or sign up for Devpost to join the conversation.