AdvisorAI

Hubs completed
Academic Reccomendations
Recommended Courses
Start Screen
Sign in Screen

Inspiration

Our main inspiration for the project was the difficulties we faced when attempting to register for courses. We noticed that even though there were many courses for electives and hub units, there was no effective way to match which hub courses best fulfill your missing units or which elective courses assisted one in working toward their desired career. Thus we sought to create AdvisorAI as a method of helping students of all years help plan and prepare themselves for the rest of their college career and beyond.

What it does

AdvisorAI uses terrierGPT, OpenAI’s API, and a database of all BU courses to pair the information of the student, BU’s courses, and future possible careers based on the students information to create a targeted roadmap of courses aimed toward preparing the student for that career. Furthermore, as it has access to terrierGPT, the program can be used for any student at any level to better understand what their goals should be and how to attain them. After it completes this task it then uses the Bureau of Labor Statistics’ database to display possible careers, their pay, and a description of the job.

How we built it

We built AdvisorAI in 4 steps. The first step was creating the AI agents in terrierGPT and chatGPT for us to get student information and parse through the Bureau of Labor Statistics’ database. Then we needed to create another step powered by AI in order to take the information and interpret it, which we created a chatGPT agent for and modified it for our purposes. After that we prompt the model that has interpreted the data to understand the current academic standing of the student and query for the necessary information and courses. Finally our frontend takes all of this information, organizes it, and displays it for the user to see.

Challenges we ran into

TerrierGPT’s performance was inconsistent, while it produced accurate results about half of the time for the primary user, any reliability quickly disappeared when the AI agent was accessed from another device. Additionally, TerrierGPT tended to hallucinate responses very quickly, often generating incorrect or irrelevant information after a few prompts. It struggled with forward thinking tasks,including reasoning through multi-step processes or maintaining logical consistency across longer interactions. These issues made it difficult to depend on TerrierGPT for generating accurate and reliable results necessary to effectively match job skills with the skills taught in BU courses. We also had several issues when dealing with the lack of an API, as incorporating the outputs of terrier GPT had to be done manually. Another challenge we had was dealing with the data. The BU courses’ database we have is mostly sanitized and processed, however we weren’t sure how to deal with the inconsistency of courses being available from semester to semester.

Accomplishments that we're proud of

Our team is most proud of the multi-agent chaining system we built. It enabled us to take a student’s input and pass it through several layers of interpretation, each agent refining and analyzing the data before producing a personalized result. This chaining process made our recommendations more context-aware and accurate. We’re especially proud of how we designed the agents within TerrierGPT to serve as the foundation, handling user data, interpreting academic goals, and retrieving relevant information from our databases with impressive precision. Getting that system to work smoothly and return reliable outputs was one of the most rewarding parts of the project.

What we learned

DS + X Hackathon’s BU Hack track specifically was a great opportunity to actually work with data and gain practical programming knowledge. We learned how to create, prompt, manage, and chain agents in terrierGPT. We also learned how to implement various AI API calls in order to better interpret and parse through data. However, beyond the technical aspects, we learned to work together efficiently under pressure and time constraints. This experience gave us insight on the numerous possibilities for incorporating AI and data science to solve real world problems.

What's next for AdvisorAI

We plan to hopefully fully integrate terrierGPT into our project to allow students to gain full access to a chain of models that have the context necessary of the student’s academic situation. We also hope to rework our integration of the chatGPT agents and API in order to improve accuracy and correctness when it comes to selecting relevant courses for the 4 year schedule and for the major specific courses. Shiboleth is another implementation we were unable to work with as we do not have the clearance to do so, however in the future by using it we can properly link a student to terrierGPT so it can query for student information.