Swiss startup Logicstar He is in love with joining AI Agent. The start-up broken in the summer of 2024 laid $ 3 million in front of the sown tools on the programmers market that may perform autonomous application maintenance, and no more typical use of the AI agent of the code coefficient coefficient.
The General Director and co-founder of Logicstar Boris Paskalev suggests that AI startup agents can end cooperation with code development agents-for example, how, let’s say, like Devin Cognition Labs-in the industry.
CODE FIDELITY is a problem for AI Agents software building and implementation, as for human programmers, and Logicstar desires to do all the pieces to lubricate the programming wheel, mechanically collecting and setting mistakes wherever they’ll appear in the implemented code.
To the present extent, Paskalev suggests that “even the best models and agents” are not in a position to solve most of the errors that are presented – hence the team spied on the opportunity to begin AI, which is involved in improving these opportunities and implementation. A dream of less tedious application maintenance.
For this purpose, they build large language models (LLM)-like Openai GPT, and even Deepeks in China-by using a model-enbrotted approach to your platform. This allows Logicstar to immerse in various LLM and maximize the usefulness of AI agents, based on which fundamental model works best to unravel a specific code problem.
Paskalev claims that the founder team has technical and specific to the domain for building a platform that may solve programming problems that may challenge LLM loneliness. They also have earlier success of entrepreneurship: he sold his earlier startup of the code review, Deepcode, and cyber security giant in September 2020.
“At the beginning we were thinking about building a large language model for the code,” said Techcrunch. “Then we realized that this will quickly become the goods … We are now building, assuming that all these large language models are there. Assuming that she is actually decent [AI] Code agents, how do we separate the maximum business value from them? “
He said that the idea was based on understanding the team, tips on how to analyze applications. “Combine it with large language models – and then focus on grounding and verifying what these large language models and AI agent suggest.”
Development based on tests
What does this mean in practice? Paskalev says that LogicStar conducts an evaluation of every application on which its technology is implemented – using “classic computer science methods” – to build a “knowledge base”. This gives an AI agent a comprehensive map of inputs and exit the software; how variables mix with functions; and all other connections and dependencies, etc.
Then, for each error presented, the AI agent is in a position to determine which parts of the application influence – enabling Logicstar to narrow the functions requiring simulation to check the results of potential corrections.
According to Paskalev, this “minimized workmanship” allows AI agent to launch “thousands” tests aimed at reproducing errors to discover the “failure test” and a-after is an approach “based on tests”-it finally ends up on the solution to repair it. sticks.
He confirms that the actual error corrections come from LLM. But because the LogicStar platform allows “very fast executive environment” that AI agents can operate on a scale to separate wheat from the chaff, and serve their users abbreviated to the best that LLM can offer.
“We see it [LLMs are] Ideal for prototyping, testing things, etc., but this is absolutely not great [code] Production, commercial applications. I think we are far from there and this is what our platform provides – he argued. “To find a way to separate these models today, we are able to actually safely separate the industrial value and save time to actually focus on essential things.”
Enterprises are set to the initial purpose of Logicstar. His “silicon agents” are to be intended to work with corporate development teams, although for a fraction of the salary required to employ a human developer, dealing with a variety of tasks related to maintaining application and releasing engineering talents for more creative and/or difficult work. (Or, at least until LLM and AI agents are much more talented.)
While the startup playground is “fully autonomous” possibilities of application maintenance, Paskalev confirms that the platform will allow programmers to review (and in one other technique to supervise), fix that AI agents call. So trust may be – and have to be – gained first.
“The accuracy provided by the human programmer is from 80 to 90%. Our goal [for our AI agents] It is alleged to be exactly there – he adds.
These are still early days for Logicstar: the version of its alpha technology is about testing with many undisclosed corporations, which Paskalev calls “project partners”. Currently, technology serves only Python – but the extensions of the typescript, JavaScript and Java are settled as “soon”.
“The main goal [with the pre-seed funding] He is actually to show that technology cooperates with our design partners – focusing on Python, “adds Paskalev. “We have already spent this year and we have many opportunities to expand. And that’s why we try to focus it first to show the value in one case. “
The pre -raised startup increase was run by the European company VC Northzone, and investors Angel from Deepmind, Fleet, Scouts, Snyk and Spotify also join the round.
In the statement Michiel Kotting, a partner in Northzone, said: “Generating a code based on AI is still at an early stage, but the increase in performance we see is already revolutionary. The potential of this technology to improve development processes, reduce costs and accelerate innovation is huge. and the huge technical knowledge of the team and documented achievements to ensure real, influential results. The future of software creation is transformed and Logicstar will play a key role in software maintenance. “
Logicstar works Waiting list For potential customers who want to specific interest in obtaining early access. He told us that the beta version is planned for later this yr.