BirdBook | Devpost

Inspiration

An estimated 11 to 36 million birds are killed illegally each year according to BirdLife International. 1,400 bird species are threatened with extinction, which 13% of all bird species We want to help researchers and conservationists monitor endangered species, help educate hunters and others about hunting season for different birds, and want to make a sustainable effort to genuinely educate about wildlife.

What it does

BirdBook allows you to input a picture of any bird and receive the name of the bird, how accurate the estimate is, and additional information regarding the bird. These results are stored in “Previous Results” in which the pictures of the birds can be saved to an interactive scrapbook.

How we built it:

We utilized Convolutional Neural Networks (CNNs), made EfficientNetB0 model, built using tensorflow, python, and matplotlib, and Gemini as our backend reasoning LLM to display generative outputs based on image indexing. We utilized Tailwind CSS, Next.js, and Splice for the front-end user interface.

Challenges we ran into: Some challenges we ran into were deciding the parameters for our model.

Gemini API integration was tricky for our first time implementation Long compute time for model generation, especially given hackathon constraints

Accomplishments that we're proud of

Successful model integration via API Successful combination of CNN and LLM for information synthesis of birds Built a model without overtraining and high precision, recall, and 99% accuracy

What we learned

APIs and implementation
Model Deployment
LLMs are not as accurate to our CNN model

What's next for BirdBook

We plan to implement a migration visualizer, which is a heatmap that displays where the birds live and migrate to. We also plan to add audio input in which you can input audio files of the birds chirps/noises and receive the same information.