localFoods

front end marked up
prediction of the state of health for images of leaves
prediction of the state of health for adulterated images of leaves
prediction of the fruits
prediction of the adulterated fruits
cockroachdb cluster
GitHub Campus!
GitHub Campus!
GitHub Campus!
GitHub Campus!

Inspiration

Everywhere you go, the effect of technology is apparent. Technology seems to have seeped into every facet of the world and the farming industry is no exception. In 2017, Amazon acquired Whole Foods, exploiting their technology to gain an unfair advantage over local farmers. Given the fact that most farmers aren’t trillion dollar companies, they are left without the reach or technological prowess to compete with monopolies like Whole Foods.

What it does

We offer a simple service that allows farmers to post their items on a website. We certify the healthiness of the item with just a picture of the leaves, using our ML model as the photo shows. The farmer just needs to take a picture of the plant and its leaves and the service will automatically classify the fruit, and post it on the website if it is healthy.

How we built it

We used Fire Module that's able to squeeze and expand its representation of the image (outlined in SqueezeNet Berkeley Research (https://arxiv.org/pdf/1602.07360.pdf)) This allows us to conserve computing resources, and offer this product at a reasonable price point, allowing all farmers to access this product. We used the RMSProp optimization as we needed to achieve the saddle point faster, due to the lack of parameters. Our Model only has 83k parameters compared to the 16 million parameters found in complex models like these. We used Google Cloud Spanner to have scalability and ACID transactions. For our login page, we used Twilio SendGrid for email verification and cockroachDB to store user data.

Challenges we ran into

We had trouble reaching the saddle point of our optimization function for ML as it was a complex task, with minimal parameters. However, we were able to reach 89% accuracy on the testing dataset after tinkering around with optimization and the overall architecture of the model. Due to the time constraint, we were not able to implement a store for consumers to buy from local farms.

Accomplishments that we're proud of

We are proud that we were able to make a working product despite the fact that most of the members in our groups were forced to use a language that we were unfamiliar with.

What we learned

We learned the Squeeze net model in order to improve the machine learning model’s efficiency for lower end devices. We learned Tailwind and utilized it in our frontend. We also learned about Cloud Spanner and cockroachDB for our database. We also learned about Twilio SendGrid api.

What's next for LocalFoods

Our plan is to make the system more robust, which would include a login/signup page, the ability to compare farms based on distance from your location and popularity, which will be calculated by the ratio of average review to the number of reviews. This would help in reducing city grid lock problems for people who live in cities and also help in net zero emissions by 2030.