Clark Smart Assistant

Campus-wide knowledge accessibility anytime, anywhere
A creative approach to a chatbot
An example of an exchange that uses both a pre-written answer and an answer generated by Roberta (Sq-2), that utilizes previous tokens
Flexible String matching ensures efficient fetching of accurate, pre-written responses

Project Aim: Clark Smart Assistant solves a problem that every student faces, a lack of accessibility to university-specific information. The goal of this project is to offer students a constant companion to help them navigate the various questions they may face during their time at Clark.
Solution: Clark Smart Assistant uses a flexible query system to fetch accurate curated answers to important questions and leverages large language models to provide additional, human-like interaction to help students solve more complex queries.
Real-World Application: The primary audience for the program is Clark students but its features could be extended to staff and faculty.
Future Aims and Goals: An additional goal for Clark's smart assistant is to offer a framework to support student extensions. By using token recognition other developers can add functionality such as finding efficient paths to different classes or fetching real-time info.

Inspiration

The design takes inspiration from smart query systems like Apple's "Siri" and question-answering ML models such as OpenAI's "ChatGPT". The idea behind the project was inspired by my own struggles to figure things out while I was a freshman.

How we built it

The project is primarily hypothetical but the example UI was built using React, and the backend is done with a mix of Python for handling machine learning and Rust for low-level computations.

Challenges we ran into

It's not really possible to fine-tune an LLM with clark-specific context in a mere 48 hours, so that integration is primarily hypothetical, however, the current framework fully supports the usage of the kind of model where it to exist today.

What's next for Clark Smart Assistant

using an NLP approach to string matching can reduce the computational demand of fuzzy string matching, a 0(n^2) operation. Additional improvements could be made using an early failure mechanism or by parallelizing the calculation of the Levenshtein distance matrix.