FUSE - Full Stack Software Engine

Describe your backend application in natural language
Sit back and relax until the code is generated and deployed for you
Check out the source code and test the deployed code

Inspiration

This project was inspired by our day-to-day experience as developers. We are already interacting a lot with tools that suggest code or are generating pieces of it and we are excited by this new wave of AI-enhanced tools and platforms. We are so enthusiastic about finding new ways of improving the developing experience using LLMs and sharing it with our fellow devs all over the world.

This is how the idea of FUSE came to us - with FUSE developers are going to be able to generate web applications by just describing the functionality in natural language.

What it does

FUSE is an AI-driven platform that generates out-of-the-box, already-deployed backend code from a simple description of the desired functionality. The platform is designed to help developers write and deploy code faster by describing the application in natural language.

How we built it

We build it using LanceDB, Langchain, OpenAI API and genezio.

Under the hood, FUSE has a two-step flow: Step 1. Generates the code using LanceDB, Langchain and OpenAI API Step 2. Deploys it to a serverless infrastructure using genezio.

For the first step, FUSE uses LanceDB, Langchain and OpenAI API to generate backend code suitable to be deployed on function-as-a-service platforms.

For the second step, the generated code is deployed using genezio because it is easy to use it. It's a single command to deploy the code to the cloud. More than that genezio generates an SDK that can be used by the frontend to call the deployed code. This simplifies the development process of the web application even more.

Challenges we ran into

Output consistency is a challenge when it comes to working with LLMs. To overcome this, we improved the model by feeding it custom embeddings using LanceDB. Furthermore, we tested a lot of prompts to come up with a robust solution.

Accomplishments that we're proud of

We are proud of building a robust AI-driven application by implementing:

a complex prompt that implements a lot of the known best practices from the prompt engineering world
a retry system in case the generated code has build-time errors.

What we learned

We learned a lot about prompt engineering and how to properly interact with an LLM, especially with the GPT models from OpenAI

What's next for FUSE - Full Stack Software Engine

We want to improve the developer ecosystem as much as possible, so probably the next step for FUSE is to generate frontend code as well. This will enable the developers to generate full-stack applications by just describing their functionality.