Startup uses LLMs to understand complex code, write documentation
A 2023 startup called Driver Inc. (Austin, Texas) has raised US$8 million in seed funding to develop its AI approach to analysing software and speeding the generation of technical documentation.
The company’s platform uses proprietary analysis software and third-party large language models (LLMs) to generate customer-facing technical support documentation.
The company estimates that it can save 50 percent of the time and resource currently put into the activity. This enables significantly faster engineer technology “onboarding” and product time-to-market. Driver makes use of third-party LLMs such as OpenAI’s ChatGPT, Anthropic’s Claude or LLama from Meta.
The technology works with software code written in somewhere between 100 and 200 languages and is particularly relevant to companies working in semiconductors and electronics in the embedded sector, Adam Tilton, CEO and co-founder of Driver, told eeNews Europe.
Previously, Tilton was the co-founder of Aktive, an embedded machine learning development platform that he sold to sports shoe company Nike in 2019, and Rithmio, which he sold to Bosch Sensortec in 2017.
Google Ventures
Driver’s funding round was led by GV (Google Ventures), with participation from Y Combinator and others. The company plans to expand the engineering team, build out the product roadmap, and accelerate commercialization.
A particular benefit is in the documentation of source code, which can be scanned, analyzed and systematized for human understanding. Millions of lines of code can be documented in two hours compared to the three months it typically takes engineers.
During project development it is possible to use the same software engine to generate application notes and code examples. The analysis can human-viewable flow-charts from code bases.
“By providing a significantly faster, user-tailored understanding of complex technology, we’re empowering teams to quickly access mission-critical information, reduce errors, and accelerate product time-to-market,” said Tilton. Being able to document software code bases of millions of lines of codes quickly and consistently enables documentation and version control to be kept up-to-date.
“Not everything is in the codebase. But what we do is enable engineers and field-application engineers to spend time on the non-tedious tasks and let Driver do the repetitive tasks,” said Tilton.
“There is no such thing as perfect documentation. You need to understand the intended audience but Driver allows engineers to generate a foundational documentation quickly. Unfortunately, it is the case that documentation tends to be an afterthought and can delay product launches,” Tilton said.
Besides working with standard software languages – including various assembly languages, C, C++, Java, Python, Rust – the Driver platform works with Verilog, System Verilog and can be applied to hardware descritions; allowing for human-understanding of complex hardware.
Driver for drivers?
Tilton said that the Driver software is capable of code generation in well-defined contexts but has not yet been applied to automated software driver generation, but that may be possible in the future.
AI-enable software, such as that supplied by Driver, may be an essential method of interacting with an interrogating extremely complex technical systems that have been partially created using AI systems.
Luna Schmid, Partner at GV, said: “We believe Driver is a game changer for any team that needs to document complicated technology quickly and ensure it can be understood by all constituents.”
The Driver Platform is usually provided under an off-premises software-as-a-service contract.
It’s features include:
- Decode vendor documentation: Reduces complex manuals into concise explanations.
- Automated updates: Syncs codebases including connection to GitHub or other source code management tools.
- Language Specialization: Works for any programming language, with specialized content for common languages and their nuances.
- Unified Search: Unified search across assets, content, and contextual information.
- Reusable templates: Create reusable templates with defined sections and instructions for how Driver should complete the document.
- Enterprise-grade security: As well as state-of-the-art encryption and advanced identity management.
Related links and articles:
News articles:
CEO interview: Machine learning boost for analog simulation
Thalia takes machine learning in EDA tool down to 22nm FDSOI