A new AI solution is capable of generating the code for complex backend software in JavaScript. This means frontend and web developers could leverage AI and natural language to create the backend code for their projects without a backend developer.
Shreds AI claims to be the first AI service to market that’s capable of generating complex software, although founder and CEO Soufiane Amar did acknowledge that Devin AI is a potential competitor. Devin AI is currently in early access.
“Shreds AI is actually the first AI capable of generating complex software, and when I say complex software, I’m talking about the software that [is] used in healthcare, finance, retail, manufacturing, energy, communication, et cetera,” Amar said. “These software are very hard to develop, cost a lot. They take months, sometimes years, and you need teams of engineers to build them.”
A screenshot of the Shreds AI developer screen.
The company is also working on a feature that will convert legacy Cobalt code to Java. It can already convert Java code to JavaScript, a process Amar called code-switching.
The solution launched in beta on August 1, and is limiting the number of customers it takes until it can update its server infrastructure to support them.
Shreds AI works by using natural language to describe what the application should do. It then generates architectural diagrams and the code for independent and isolated features called “shreds.”
“These softwares are very hard to develop, cost a lot. They take months, sometimes years, and you need teams of engineers to to build them.”
– Shreds AI founder and CEO Soufiane Amar
So far, it is an “expert” in Java and all related frameworks, including Spring/Spring Boot, as well as Java tools, Amar said. It’s also “good” at JavaScript code, with support for React, although that is not as comprehensive as its support for Spring, he added. The team is training the models and plans to open it up for Python, PHP and 30 other languages and frameworks — including Laravel and Django — in the next 4 to 6 weeks.
For anyone doubting this, you’re invited to look at a GitHub repository that holds Java 17 code Shreds generated to build a WordPress clone based on a microservices architecture. The Shreds AI team created it without access to the WordPress code, instead relying solely on API documentation, according to the repository notes.
The Elephant in the Application
At this point, the AI is not trained to build user interfaces, although Amar said that is on the roadmap. Currently, it’s restricted to creating backend code, which can be in JavaScript and Java.
Recognizing that this might not be the most welcome of news to developers (they’re working on training the AI to write complex articles so, believe me, I get it), I asked Amar what this means for developers.
The tool still requires human developers to review the code, he said — which assumes developers would welcome being relegated to doing code reviews for AI. In fact, code review is built into the platform: It includes a marketplace of freelance developers who will check the code for you, although the system also supports internal code review by on-staff developers.
This is how the platform ensures there are no hallucinations or incorrect code, which is a known issue for AI. It sounds time-consuming, but Amar insisted it’s not.
“It takes only a few minutes, or like 30 minutes, to validate the code, because of its precision in about 93 to 98% in the code,” he said. “It’s a quick process, and you get your code generated, validated by human beings and ready for production.”
It’s not hard to see the business case. Amar estimated the cost of an average software application using Shreds would cost around $23,000 compared to about $700,000 for a regular engineering team.
How Shreds.AI Works
The AI is specifically trained to manage software complexity, but it is also trained to use other available AIs to solve complex problems, he explained. It functions like a Google for AI, he said — it ranks the models, identifies weaknesses and strengths, and then uses the appropriate model to overcome hard programming problems, he said.
“For example, if you need to use a math formula for calculating lifespan average or something, then Shreds AI can make multiple models, including GPT or Gemini or other popular models, to build that solution, then integrate it into the software,” he said. “This is how Shreds AI is capable of literally solving any programming problem that it comes across.”
Graph showing how Shreds AI leverages other AI models.
It “makes” the models collaborate, he added, aggregating the intelligence and using it to solve complex problems.
“It has an algorithm that allows it to go to that model and test its strengths and weaknesses, and it knows whether it’s good at math, whether it’s good at logic, whether it’s good at writing, at different things,” he said. “We can use the strengths and weaknesses of tens of models to solve complex problems and produce some consistent and high-value results.”
The code is generated in pieces called “shreds,” which is a workflow in your software. The shred is then pushed to the marketplace for review, he said.
Currently, the company is developing the feature to convert legacy Cobalt code to new Java code. Many financial institutions and government applications have been built in Cobalt, and updating that code is difficult because there are fewer Cobalt programmers now. The plan is to add other languages in the future, he said.
Early Testers
Shreds AI is available in the cloud, but it’s also possible for large companies to run it on-premise.
So far, the beta testers include large companies such as Samsung, the car company Stellantis, and French electric company RTE. For Stellantis, Shreds AI built a system that allows car owners to open their cars using their smartphones. RTE used it to generate a map that maps the live electricity production and consumption from nuclear plants.
The post Shreds AI Creates Complex Software in Java and JavaScript appeared first on The New Stack.
The solution currently produces Java, JavaScript and React, with plans to expand to 30 other languages and frameworks within two months.