Sweep's AI Powered Junior Developer

Though the world of software development seems undoubtedly dynamic, with more and more useful and exciting applications and websites we see adding up every day, it is also absolutely true that developers find themselves bogged down by mundane but time-consuming and resource-draining tasks.

Such tasks hinder developers from focusing on more challenging and impactful projects. According to Stack OverFlow’s 2023 Developer Survey, which had over 90,000 responses from developers, 63% of 50-teamed developers lose between 333 to 651 hours per week searching for answers or solutions to problems, which can include bug fixes, code refactoring, or edge cases.

Sweep, a YC-backed startup founded by William Zeng and Kevin Lu, who both worked at Roblox and found the issue of repetitive, mundane tasks especially apparent, creates an AI-powered junior developer platform that can handle tech debt and debug with pull requests.

AI’s Job

The necessary but frustrating stack of mundane tasks developers have to deal with every day diverts their attention from actual high-skilled coding that supports their creative endeavors.

AI chatbots and large language models have become one of the solutions for developers to perform mundane coding tasks that can very much drain their day-to-day productivity. However, LLMs have the tendency to underperform when it comes to coding, getting things wrong on their first try. If no rechecks are done, developers would have to spend twice as much time fixing the LLM’s mistakes.

Other ways to automate such tasks have been developed by several companies. AI-powered platforms like CodeRobot and DeepCode help make a developer’s process more efficient and productive with their automated code generation, code testing, and code review features.

Sweep’s co-founders knew that the software chores that they had to deal with while working at Roblox could be automated with AI, the tech just needed a little push. This is where the Sweep platform comes in.

Sweep’s Magic

Sweep has the ability to write, fix, and debug its own pull requests without any need for human intervention, saving developers from having to waste time and effort manually performing these tasks.

Using machine learning and natural language processing, Sweep can understand the code that exists in a codebase, identify potential problems, and generate the code changes necessary to fix said problems. What this means is that developers can use Sweep to describe their coding requests in plain English without any need to write the code themselves. Sweep can understand the request and be the one generating the corresponding code.

Since the platform is intended to act as if it were an AI-powered, meaning fully autonomous, junior developer, Sweep is also designed to push the generated code to the appropriate codebase and address any comments made by senior developers on the pull request. Sweep can also fix its own pull requests on any GitHub lints and failed tests or builds.

Python Dealings

Sweep specializes in generating Python code by utilizing various AI models like GPT4 for code generation, but it also uses a custom code search engine that allows Sweep to execute code changes affecting an entire code repository. The co-founders assert that the engine isn’t trained on customer data to prevent any biases towards any particular company or project. The code search engine runs on lexical and vector search techniques.

To further enhance its effectiveness, Sweep is given the ability to read and write large files of over 2,000 lines of code quite quickly. When it comes to gargantuan tasks, however, Sweep also has a slow mode feature that allows it to take extra time to carefully execute said tasks.

Developers can set Sweep up by simply adding the Sweep GitHub app to desired repositories and creating new issues in the repo. Simpler issues can be solved by Sweep with a single line and a file name, while harder problems would require developers to give Sweep more information like they would a junior developer.

Sweep raised $2M from Goat Capital, Replit CEO Adjad Masad, Replit VP of AI Michele Catasta, and Exceptional Capital. The startup aims to use the funds to extend its team from two employees to five and further improve Sweep’s features.

Meme & AI-Generated Picture

Job Posting

• Epam Systems - DevOps Engineer - New York City, NY (Hybrid)

• GrubHub - Sr. Staff Engineer - New York City, NY (Hybrid)

• Roofr - Machine Learning Engineer - US (Remote)

• Upgrade, Inc. - Senior Analyst, Capital Markets Operations - San Francisco, CA (Hybrid)

Promote your product/service to Digger Insights’ Community

Advertise with Digger Insights. Digger Insights’ Miners are professionals and business owners with diverse Industry backgrounds who are looking for interesting and helpful tools, products, services, jobs, events, apps, and books. Email us [email protected]

Your feedback would be greatly appreciated. Send it to [email protected]