RSHK | Devpost

Screenshot of the homepage

Inspiration

We were messing around and coming up with some weird ideas, this one happened to be pretty good

What it does

The web applet accepts user-specified geolocation data from a graphical interface as well as a magnitude of earthquake and, using a trained neural network, outputs the estimated total injuries caused by the earthquake and the estimated cost in infrastructural damages.

How we built it

We trained a neural net through ibm-watson on data retrieved from the noaa on earthquakes going as far back as 1973, then using flask and jquery set up a web applet to accept an input for the neural net, which then spits out the output.

Challenges we ran into

Documentation for most elements of this project (census bureau data, ibm-watson, etc) were sparse and/or inconsistent, and we were all stepping out of our comfort zones in making this, so we ran into issues in nearly every step of the project. Not to mention, the flask server was finnicky and randomly either broke or failed to connect to our backend (this too was a consequence of sparse documentation), and also in general managing the AI in terms of how it should read the data -- it took a few tries before watson fully understood that the magnitude of the earthquake affects the damage it causes -- and the initial data set was imperfect and required pruning to get into a shape ready for the AI.

Accomplishments that we're proud of / what we learned

None of us on the team were experienced at all in anything we used in this project: ibm-watson, flask, web development, and this project, all things considered, was a success. We all learned a great deal about all these, which are no doubt valuable skills to have going forward.

What's next for RSHK

RSHK as it stands is somewhat crude; currently the way it tracks population given a location is to use the population data of the county surrounding it. It's a fine heuristic, but it means that for example a magnitude 8 earthquake will affect the exact same amount of people as a magnitude 3 earthquake, which is not entirely accurate. Moving forward it would be better for the population to be the amount of people in a radius dependent on the magnitude from the initial location. Also, there is always room for more data to train the AI on; more data will make the neural network more accurate, which is always better.