SIMPL-AI

The homepage, with an entry bar to ask Gemini for some help!
The network builder page, where users can configure network parameters.
The training page, where users can upload CSV data to train their network.
The prediction page, where users can run predictions with their trained model.

Inspiration

When you first hear of the phrases neural networks and reinforcement learning, often you are left with a feeling of dread, because although they sound like incredibly cool topics, they sound very daunting and scary to dive into. Mirroring our own experiences, we also experienced this barrier, and indeed, learning neural networks was fairly challenging because it was hard to understand how changing various parameters would affect the performance of the neural network. Thus, we decided to create SIMPL-AI.

What it does

SIMPL-AI is a neural network builder that provides a simple tool for developers looking for an easy way to tweak neural network parameters in a nice interface, or for students who are first encountering neural networks to easily explore and experiment with various configurations. Our tool emphasizes education and equity/inclusion by providing a simple and easy-to-use pipeline from network specification to prediction generation. On our home page, we also include a space for individuals to type in a problem that they would like the neural network to tackle. This provides further convenience for developers and students to easily create a neural network and explore, experiment, and tweak their neural network without having to take on the complex code.

How we built it

We built this application with a combination of web technologies and tensorflow. By leveraging the TensorflowJS library, we were able to construct a Next.js based environment where individuals are able to send in their network parameters and create and train models. Additionally, we leveraged the Google Gemini API to provide parameter setting advice to individuals who are looking to learn and may not know as much about the process.

Challenges we ran into

One of the first major roadblocks was getting TensorflowJS to run in the Next.js environment. We overcame this issue by first getting everything to work in a native Node.js environment with static inputs, and migrating it piece by piece to a version on Next that transcribed user input into the parameters. Another roadblock was getting the Google Gemini API set up, as none of us had used it before. It took a bit of reading through documentation, but eventually the Langchain setup worked out for us and we were quickly delving into some exciting prompt engineering to get the desired output from the chatbot.

Accomplishments that we're proud of

We are quite proud of integrating something as advanced as AI/ML into a web based environment of our own. We're also proud of learning how to use the various tools we did in the short time frame of this hackathon. Last but not least, we're proud of the knowledge regarding neural network architechture all of us accrued throughout this project.

What we learned

We learnt a lot throughout this project. Not only did we have to brush up on our web development skills, but for the majority of our team, we had never touched AI in this theoretical depth before. Figuring out all the intricacies behind model architecture and hyperparameter optimization was quite challenging, but also rewarding. We will definitely take forward this theoretical knowledge into future coding projects we create.

What's next for SIMPL-AI

This project can only grow into something bigger. There are many more network architectures and parameter types waiting to be incorporated into this model. Additionally, the LLM-based help is only a basic proof of concept at the moment, and could be integrated far deeper into the actual process.