The Swiss startup LogicStar, founded in summer 2024, has secured $3 million in pre-seed funding. The company aims to introduce tools for developers that facilitate the autonomous maintenance of software applications, differing from the common applications of AI agents that focus on code co-development.
Boris Paskalev, the CEO and co-founder of LogicStar (featured in the image above with his co-founders), indicates that the AI agents created by the startup may collaborate with code development agents—like Cognition AI’s Devin—to create a mutually beneficial business environment.
Code accuracy presents challenges for AI agents involved in software creation and deployment, just as it does for human programmers. LogicStar is determined to enhance the development process by automatically identifying and resolving bugs that appear in deployed code.
Currently, Paskalev points out that “even the best models and agents” cannot tackle a significant proportion of bugs presented to them. This has led the team to see an opportunity to establish an AI startup focused on boosting these chances and achieving the goal of simpler application maintenance.
To achieve this aim, LogicStar is building its tools using large language models (LLMs)—including OpenAI’s GPT and China’s DeepSeek—employing a model-agnostic strategy. This flexibility allows LogicStar to utilise various LLMs and optimise the effectiveness of its AI agents based on which foundational model is best suited for addressing specific coding issues.
Paskalev asserts that the founding team possesses the necessary technical acumen and industry-specific insight to create a platform capable of resolving programming challenges that can perplex or outsmart LLMs working in isolation. Additionally, the team brings with them a record of past entrepreneurial success, as he previously sold his code review startup, DeepCode, to the cybersecurity firm Snyk in September 2020.
In the initial stages, the intention was to construct a large language model specifically for coding, Paskalev stated. However, they soon realised such models would quickly become standardised. Consequently, the focus shifted to leveraging existing large language models and AI agents to extract maximum business value from their capabilities.
He added that their concept is rooted in their comprehension of software application analysis. By merging that knowledge with large language models, they aim to ensure verification and grounding of the recommendations made by those models and AI agents.
Test-driven development
What does this look like in practice? According to Paskalev, LogicStar conducts an application analysis using “traditional computer science techniques” to create a “knowledge base.” This comprehensive framework allows the AI agent to understand the software’s inputs and outputs, relationships between variables and functions, and other connections and dependencies.
For each bug encountered, the AI agent can identify which components of the application are affected, enabling LogicStar to focus on simulating the necessary functions to test numerous potential fixes.
Paskalev explains that this “minimised execution environment” allows the AI agent to perform “thousands” of tests to reproduce bugs and find a “failing test.” Through this approach known as “test-driven development,” the intention is to arrive at a lasting solution.
The bug fixes themselves are derived from the LLMs. However, as the platform allows for a “very fast executive environment,” LogicStar’s AI agents can effectively sift through options, offering users an efficient means to access the best solutions available from LLMs.
Paskalev noted that while LLMs excel at prototyping and testing, they are far from suitable for production-level commercial applications. He expressed confidence that LogicStar’s platform bridges this gap, enabling users to leverage AI effectively while saving developers time for more critical tasks.
The primary target audience for LogicStar will be enterprises. Its “silicon agents” are designed to support corporate development teams at a significantly lower cost than employing a human developer, managing various application maintenance duties and freeing engineers to focus on creative or complex challenges. This is expected to be the case until AI agents and LLMs advance further in their capabilities.
Although the startup promotes its technology as offering “fully autonomous” app maintenance, Paskalev confirms that the platform enables human developers to review and supervise the fixes proposed by its AI agents, indicating the importance of establishing trust.
The accuracy achieved by human developers ranges between 80 to 90%. The target set for their AI agents is to match that level.
It is still in the evolving phase for LogicStar. An alpha version of its solution is currently being tested with several unnamed companies, referred to by Paskalev as “design partners.” Currently, the technology is limited to Python, but plans are in place to expand to TypeScript, JavaScript, and Java soon.
Paskalev mentioned that the pre-seed funding aims to demonstrate the technology’s functionality alongside their design partners, initially concentrating on Python. They have already invested a year into this development and see ample opportunity for expansion, which is why they are focusing initially on illustrating its value in a singular case.
The startup’s pre-seed funding round was led by Northzone, a European VC firm, with contributions from angel investors associated with DeepMind, Fleet, Sequoia scouts, Snyk, and Spotify.
In a statement, Michiel Kotting, a partner at Northzone, remarked: “The early stages of AI-driven code generation have produced revolutionary productivity gains. This technology holds immense promise to streamline development processes, reduce expenses, and foster innovation. The team’s extensive technical expertise and proven track record position them to make a significant impact. The landscape of software development is evolving, and LogicStar is poised to be a key player in software maintenance.”
LogicStar is currently maintaining a waiting list for interested customers seeking early access. A beta release is anticipated for later this year.
