LLM-Powered Workflow Optimization for Multidisciplinary Software Development: An Automotive Industry Case Study
This paper presents a graph-based, LLM-powered workflow optimization approach that bridges the gap between domain experts and developers in multidisciplinary automotive software development, achieving a 93.7% F1 score and reducing per-API development time from 5 hours to under 7 minutes in a production Volvo system.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are building a massive, high-tech car. You have two groups of people trying to work together:
- The Experts: These are the engineers who understand the car's physics, the electrical signals, and the safety rules. They speak a language of "CAN signals," "voltage ranges," and "brake pressure."
- The Developers: These are the coders who build the apps you see on your phone or the dashboard screen. They speak a language of "APIs," "JSON," and "Python code."
The Problem: The "Lost in Translation" Bottleneck
In the old way of doing things (described in the paper), these two groups were like two people trying to build a house while speaking different languages and standing in separate rooms.
The Experts would write a manual saying, "The AC should be 'Economy' mode when the signal is 0."
The Developers would have to read that manual, figure out what "0" means, find the right wire, and then write code to make the AC switch.
This process was a nightmare because:
- It was slow: It took about 5 hours to write just one small feature.
- It was messy: If the Experts changed their mind, they had to tell the Developers, who had to stop and rewrite everything.
- It was error-prone: Humans get tired. A developer might misread "0" as "1," and suddenly your car's AC blows hot air when it should be cold.
The Solution: The "Super-Translator" Robot
The researchers at Volvo and Chalmers University built a smart system using Large Language Models (LLMs)—the same kind of AI behind tools like ChatGPT—to act as a Super-Translator.
Think of this system not just as a "spellchecker" for code, but as a construction foreman that sits between the Experts and the Developers.
Here is how the new system works, using a simple analogy:
1. The Blueprint (The Graph)
First, the team mapped out the entire construction process like a subway map. They drew lines showing who talks to whom and what information needs to move. They realized the map was too crowded with "human stops" (meetings, emails, clarifications).
2. The Three-Step Assembly Line
Instead of humans doing the heavy lifting, they built an automated assembly line with three stations, each run by a specialized AI robot:
- Station 1: The Signal Decoder.
- Old way: A human reads a technical manual about a car signal and manually types out code to read it.
- New way: The AI reads the manual and instantly writes the code to "listen" to that signal. It's like a robot that instantly translates a foreign recipe into a shopping list.
- Station 2: The Matchmaker.
- Old way: A human has to guess which car signal controls the "Climate Mode" button on the app. They might get it wrong and have to ask the Expert, "Hey, does this button mean Economy or Standard?"
- New way: The AI looks at the app button and the car signals, figures out the connection, and writes a rule: "If the app says 'Economy,' set the signal to 0." It does this with high precision, checking its own work.
- Station 3: The Builder.
- Old way: A human takes all those rules and assembles them into a final app feature, making sure the formatting is perfect.
- New way: The AI takes the rules and snaps them into a perfect, ready-to-use app feature, just like a 3D printer building a part from a digital file.
The Results: From Hours to Minutes
The results of this "Super-Translator" system were incredible:
- Speed: What used to take 5 hours per feature now takes less than 7 minutes. That's like going from walking across a city to taking a high-speed train.
- Quality: The AI got it right 93.7% of the time on the first try. The remaining errors were caught by the system's self-checking mechanism (it writes its own test questions to see if it passed).
- Happiness: Both the Experts and the Developers were happier. The Experts didn't have to spend all day answering emails, and the Developers didn't have to spend all day guessing what the Experts meant. Everyone reported high satisfaction.
The Big Picture
This paper isn't just about writing code faster. It's about changing how teams talk to each other.
In the past, the "translation" between different types of experts was the bottleneck. By using AI to handle the boring, repetitive translation work, humans are freed up to do what they do best: thinking, designing, and solving hard problems.
It's like giving a construction crew a team of robots that handle all the brick-laying and measuring. The human architects and engineers can then focus on making sure the building is beautiful, safe, and functional, rather than worrying about whether they counted the bricks correctly.
In short: They turned a slow, error-prone game of "Telephone" between different experts into a fast, automated assembly line, saving nearly 1,000 hours of work and making the car software much more reliable.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.