What it does
Sleep Search answers your medical questions about sleep disorders! Just as a question in a natural language like "how do I treat insomnia?" -- and it will give you a helpful answer without needing to surf through endless articles and ads like you'd have to with Google.
Why we built this
Inspired by the hackathon name, and growing & major issue in the us, people with sleep disorders have a hard time improving their sleep quality and getting answers to their tough medical questions. So we set out to solve this problem this weekend! Watch another video we created talking about what inspired us here: https://www.loom.com/share/4c74eefb0fe846b49d5e9ec2f99175b8
How it works
this app contain 5 layers of NLP machine learning, NLP deep learning, and a web scrape api.
- first we have naive bayes model that determine where should our question passed to
- second we have a naive bayes model again, to understand nlp input from user, and search our database for the answer. btw our database can contain up to 10,000 data. We use google cloud to host the mongodb
- third, we have deep learning gpt3 model trained in thousand of data, to answer the question if the naive bayes model not quite sure of the answer.
- fourth, we have wikipedia api and google web scrape api for the backup plan if all our trained model not quite sure of the answer.
- and last, we have wolfram api for the backup plan if the question cant be answered on previous layers.
but 90% of the question can be answered by our model 1st until 3rd layer, so the 4th and 5th model is just for a backup plan if our model still not trained on that data
How we built it
First, Arnav, Evan, and Dominick working on the front end to make some nice and simple web interface, meanwhile Hazel start to scrape some data to train naive bayes model on google colab fun fact : we use naive bayes model for nlp bcs it has the best performance with low resource use, meanwhile the tensorflow model is so heavy as well as slow that it cant be deployed on heroku. Next Arnav and Hazel start to built the backend flask app and connect it to mongodb on google cloud, meanwhile Dominick start to tweak GPT3 a pretrained model. After that we all work together to connect all the model and the frontend. Then we deploy all of it on heroku
Challenges we ran into
- scraping the data (solved)
- integrating all layers (solved)
- integrating frontend to backend (solved)
Accomplishments that we're proud of
We were proud of being able to make a project with this many layers, so much data, and so many bugs, fixed and organized in time to submit to the hackathon, making this project was really fun.
What we learned
- working as a team
- learning new framework like MongoDB
- learning new library like sklearn, pickle, and pymongo
- learning the fundamentals of artificial intelligence and machine learning
What's next for Sleep Search
- We will train our model periodically so the model will always be up to date
- We will add more answers to our database
if you want to search for the second time, you have to go back to https://sleepsearch.herokuapp.com/ (the link without any query in it) so that the app can work properly again ITS NOT A BUG, ITS A FEATURE (FOR SOME SECURITY REASON)