Nebo

Inspiration

Agriculture is one of the major sources of income in 3rd world countries. In the age of AI we have seen problems like self driving take first priorities and other really cool research, but I haven't seen a good enough solution that helps the average farmer get better yields, understand what is really the problem with their crop. I haven't seen other startups have access to the power that deep learning can have in order to integrate the same into their products and help solve even bigger, broader scope of problems a single person like me can't solve or don't know much about.

E.g a food supply startup can leverage the treatment of a certain pathogen and supply pesticides and back to the farmer while still getting farm produce from them

What it does

Uses object detection to detect plant pathogens and gpt for answering questions about the pathogen, from what causes the said pathogen, ways of treatment and any other information related to the said pathogen.

How we built it

The project is built using a combination of multiple technologies, the backend uses fastapi, tge question answering model currently in use is GPT3 davinci model but switching to a self hosted question answering model since the current cost of gpt3 is quite high for the current usecase so after a bit of research and fine tuning we are going to switch to a model from hugging face. The object detection model is built using yolov5 from ultralytics. I found it to be pretty straightforward to do training and learnt a lot about localisation.

Challenges we ran into

The main challenge was the lack of available annotated agricultural data. This meant that I had to learn everything I could about how the pathogen looks like then collect images and manually annotate them.

The annotation process took the most time and currently the model still needs more data. This is something the help of an agronomer would have helped a lot.

The other challenge came from having to learn about android development managing states etc

Accomplishments that we're proud of

Learning a lot about localisation and getting to a fully deployed backend using docker in swarm mode.

Seeing all this work is pretty much the one major thing am proud of

What we learned

Got to learn about on device model deployment (not included in the current app). How image localisation works. How to effeciently deploy a PyTorch model in production and how to version different models, adding to automatically pushing the new model to production and monitoring data drift using weights and biases.

What's next for Nebo

Switch to a more B2B model, where we provide the API endpoints as services other agricultural startups can use including agricultural research institutes. This will help with a wider reach of the technology since most of the customers of the startups are already farmers. Leveraging that reach by providing our endpoints for them to easily integrate into their services will help it have a much larger impact.

Providing the service as a WhatsApp chatbot, reason for this is most farmers (after doing some research) already have WhatsApp installed on their phones, this helps by reducing the friction of requiring a farmer install another app, plus cuts down on development time, by using the WhatsApp platform we get to have all the advantages that comes with it.

A proper UI/UX design of a simple android application (for people who really want to use it and learn about the diseases without using the chatbot). Once their's a design build and push the app to play store.

But main focus is providing these services to already existing startups who don't have a budget to hire machine learning engineers or are slow to adopt the technology due to some other reasons

Built With

• dart
• fastapi
• gcp
• postgresql
• python
• pytorch
• yolov5