-
-
Interacting with the model.
Inspiration 💡
We were inspired to create this because we wanted to redefine the user experience of the Verizon website for customers.
What it does 🤖
Verizon Vera is a virtual assistant that takes input about customer needs and recommends real Verizon apps and services to them.
How we built it 🛠️
We built Verizon Vera by using a local LLM trained by Mistral AI. We fine-tuned this model using real Verizon product data stored within its context.
Challenges we ran into 🙅
One of the challenges for building this was finding the correct parameters for optimal performance. Too much to the side of speed, and the model will lose context of the product data. Too much to the side of information, and the model will take too long to respond.
Another challenge we ran into was using voice recognition as input. It worked on Windows, but not other OSs. The difference between each team member's OS and architecture lead to difficulties with running all the software on one machine. We had a similar issue with text-to-speech for reading back LLM responses to the user.
Accomplishments that we're proud of 🏆
We are proud that were were able to use a local LLM, as opposed to a hosted one, like ChatGPT. We are also proud that the model is able to accurately recommend products and even give follow-up information on things like price and details.
What we learned 🧠
We learned how to use local models, how to fine-tune LLMs, how to do transfer learning, and how to integrate all of that into the website.
What's next for Verizon Vera 🔮
The next step of Verizon Vera is to allow voice input for Mac & Linux. We could further improve performance of the model, giving it faster responses. This would be done by further optimizing the model parameters and running it on more robust hardware.
Team Members:
Gowthami Wudaru - Computer Science
Ansari Shaik - Agronomy
Daniel Gruen - Computer Science
Tyler Woodruff - Computer Science
Log in or sign up for Devpost to join the conversation.