A startup called Imandra Inc. says it’s taking artificial intelligence-driven code completion to the next level with the launch of an entirely new and automated reasoning system called CodeLogician.

Unlike other large language models designed specifically for code completion tasks, such as GitHub Inc.’s Copilot, CodeLogician is said to be based on a new concept called “neurosymbolic AI,” which allows it to apply reasoning on the code it generates, making it far less susceptible to so-called “hallucinations” or inaccuracies.

CodeLogician is powered by ImandraX, the latest version of the Imandra Core reasoning engine that’s widely used in the financial services industry and government sector to verify, test and audit mission-critical systems, including national stock exchanges.

By automatically converting the code it creates into mathematical models, CodeLogician can then leverage the ImandraX engine to better understand, analyze and verify application-level source code. It can also automatically generate tests that prove the accuracy of its code. As such, CodeLogician is not just an AI coding assistant, but also a verification tool that guarantees the veracity of the code it generates, helping developers catch any security vulnerabilities and prove it functions as intended.

CodeLogician was built using the LangGraph framework and its initial release is compatible with the Python programming language, with future updates set to add support for Java and COBOL, so it will be able to help transform legacy software applications.

According to Imandra, CodeLogician will transform developer productivity, eliminating the burden of having to validate thousands of lines of AI-generated code manually.

Imandra co-founder and co-Chief Executive Grant Passmore said existing generative AI coding tools are flawed, because although they generate lots of plausible-looking code, there’s no way for them to guarantee the accuracy of that code.

“The code is often wrong in subtle and dangerous ways,” he said. “CodeLogician goes beyond generative AI, using symbolic mathematical reasoning to ensure code actually behaves as intended.”

Imandra said CodeLogician’s secret sauce is the LangGraph framework, which allows it to iteratively refine its underlying models, explain its reasoning and deliver high-assurance guarantees. To do this, it relies on a unique property known as “state-space exploration,” which allows it to analyze exhaustively all of the possible states and behaviors of an LLM using symbolic region decompositions.

It’s a novel technique that allows it to understand what the underlying LLM is thinking when it tries to solve a problem. That helps it ensure the accuracy of the code generated by the LLM, before applying intelligent tests that prove its correctness.

CodeLogician is available now for early testers and developers are invited to sign up to a waitlist for access. Once it launches, it will be made available programmatically through an application programming interface and also as a VS Code extension in Microsoft Corp.’s Visual Studio Code Marketplace.

“Bill Gates famously dubbed formal methods for general software development as the ‘Holy Grail,’” said Imandra’s other co-CEO Denis Ignatovich. “Now, with the power of neurosymbolic AI and breakthroughs in automated reasoning, we’re stepping ever closer to achieving that goal — putting advanced reasoning tools directly into the hands of engineers.”

Image: SiliconANGLE/Freepik AI Suite