Climodelity

Inspiration

We realized that many people have been trying to build or have built machine learning models to predict natural disasters but these models can experience failure, with the cost of this mistake often being human lives. We want to prevent this with our product, Climodelity. Special thanks to additional inspiration form Andrej Karpathy and his work in Auto Research that helped inspire our implementations of the concept.

What it does

Climodelity is an end to end pipeline that helps test your models. By generating hypotheses and through welch corrected t-tests, Climodelity determines how the model fails and how it can be improved, showing data analytics of the model's performance and suggestions in a sleek dashboard.

How we built it

We built it in stages, first designing the Auto Research component, while other team members worked on the front end, intermediary preprocessing layer, and the model we were going to test as demo(which was trained on one of the required datasets). We did use the assistance of claude code in the endeavor. By combining each members work into a greater whole, we successfully created Climodelity.

Challenges we ran into

One challenge we ran into was finding suitable data to train our model on, which we resolved through parsing large datasets for the required timestamps and location data we were looking for. Another challenge we ran into was determining how to objectively measure the quality of a model, which we took under great consideration and decided that the best approach was to stick with RMSE to evaluate the model.

Accomplishments that we're proud of

We are proud of solving the above challenges with great thoughtfulness and in implementing and creating this complex analysis tool to be as user-friendly as possible.

What we learned

We learned that when evaluating machine learning models there are many considerations that need to be noted, and that when properly guided, machine learning can be incredibly effective in the protection and safeguarding of human life. Furthermore, the primary consideration with any model, must be how dangerous it is for the model to be incorrect before you procced. Also, throughout the course of this hackathon, our team had to learn to wear many different hats and work incredibly efficiently to complete this massive project.

What's next for Climodelity

We want to talk to environmental scientist and people who develop these natural disaster models for user input. We want our Climodelity to make a real difference and help save lives, and we believe that talking to specialists on how they can integrate our product into their workflow is the next step in achieving our goal.