P² (PodPilot)

p^2_logo

Inspiration

Data center technicians often have an overwhelming schedule, especially due to it being unorganized and inefficient.

What it does

PodPilot enhances the work of data center technicians by optimizing their schedule taking into account the distance between their jobs, the severity of their jobs, and the skills of the technicians. By using an intuitive UI, PodPilot allows technicians to complete their work efficiently and gain insight on unseen problems through the use of generative AI.

How we built it

We first focused on the schedule optimizer, as it’s a key value proposition of our project. We wanted to use an optimizer that took various elements into consideration. To do this, we decided on using the Nvidia cuOpt model which provides routing optimization that enables users to solve complex optimization problems efficiently. This made it perfect for our use case. The backend was built out in Supabase, allowing us to easily generate synthetic data for technicians as well as tickets (representing the issues in the data center). Finally, we built out a frontend within Streamlit which provided an interactive dashboard for technicians to view their routes for the day and overall metrics, as well as allowed them to access the Nvidia llama-3_1-nemotron-ultra-253b-v1 model to gain insight on never before seen issues and even upload pictures.

Challenges we ran into

Requirements analysis stage : when we first chose this challenge, we were not aware of what the day to day of a data center technician looked like, let alone the problems. To proceed with this challenge statement, we had to perform intense requirements analysis to gain an understanding of this industry and the problems that would need to be addressed. Due to this, the ideation phase of this project took up more time than expected. New tech stack for all team members : as a team we were all new to using Streamlit. We initially chose it because we found that it would be an easier and lighter alternative than using frameworks such as React.js or Next.js. The fact that it was new to all of us caused debugging to take a lot longer than expected. Using Nvidia cuOpt : it took a lot of time to understand the structure of the inputs for cuOpt and the overall data preprocessing that it took before creating an optimized solution. Additionally, because we were using synthetic data, it took many iterations before reaching a solution that we felt was feasible.

Accomplishments that we're proud of

Achieving our MVP Having a thoroughly thought out application idea

What we learned

Day to day life of a data center technician More about Nvidia’s tools and their large network of models

What's next for P² (PodPilot)

Utilize integrations in Supabase to trigger a webhook upon the creation of a JIRA ticket and automatically update the database in Supabase. On-Site Verification: confirm correct rack/server/cable using OCR / Vision Recognition Technicians can mark days they will be out of office so that the schedule can reassign tasks. Manual override rights for admins.