Week 24: The final solution

The final solution of our project consisted in a chatbot that allows for the previous projects to be more transparent to the end users by letting them question why certain decisions were taken or by providing some feedback and make some requests regarding their energy.

The team's final thoughts:

Bruno Ribeiro - Natural Language is one of the most interesting areas related to artificial intelligence because communicating is one of the primal things humans learned to do and being able to build a software that learns that, is fascinating. It was very interesting to work with Natural Language tools and be able to see the robustness and the advanced we are in this field and also the components working behind these tools as well as their limitations.

Bruno Veiga - The project was very interesting not only to create a chatbot that could be used by the communities of our older projects but also to gain new knowledge regarding natural language understanding and the deep neural networks. Both themes are probably my favorites to work on so far and it feels like there are so many more interesting things to be explored, especially with other contents that we did not have the opportunity to explore practically in this project which is very motivating.

Carlos Coelho - The chatbot idea came as an obvious next step in our applications catalog. If we are trying to provide a service for energy management, there must be a way for the participants to clear their doubts. The whole process from thinking about how to train the models to creating the datasets and seeing the trained models in action was very interesting and actually fun, to the point we decided to create those extra models to complement our solution. In the end, it was outperformed by the GPT-3 model, but it was never about the end result, it was about the journey. (Or so they say)

Miguel Silva - The decision of developing a chatbot that would help a community interact with energy-related doubts, explanations, and suggestions, was a challenge that, on a personal note, was quite interesting, particularly the building of the dataset and the mechanics for the chatbot to learn how to identify the intents and tags, so that it would, afterward, provide with an appropriate answer. It was also quite impressive to realize how much the pre-trained models such as BERT and especially GPT-3 have evolved, as the latter was able to make these identifications with such little dataset to train on and still have amazing results.