AI for Problem Solving (At Scale)

Today’s edition is co-authored by Valentin Marguet, project lead in the automotive industry. The AI projects described here are personal experiments outside his role.

"For every problem we solve, ten new ones appear in the backlog". Sound familiar? The real power of AI isn't in fancy tech demos – it's in fixing those everyday operational headaches that drain your team's energy.

Today, I'm sharing my fourth collaboration with Valentin, who's been spending his nights for the past 2 years studying and building AI side projects for high-performance engineering leveraging its own experience at one of the most prestigious automakers in the world.

And this is where he recently discovered how AI can help to overcome a persistent industry challenge: the endless technical problems backlog.

Let's dive in!

📢 Workshop Announcement

Problems, Problems Everywhere!

Not only - but especially - in the world of automotive engineering, technical problems are as inevitable as deadlines.

Every week brings new issues: components failing in extreme conditions, materials behaving unexpectedly, complex systems showcasing unplanned behaviors…

As Valentin says "For engineering teams, this means an ever-growing backlog of technical issues waiting for investigation". He discovered: engineers weren't spending most of their time fixing problems – they were spending it trying to understand them in the first place!

One of these painful areas was conducting root cause analyses. Hours upon hours sifting through test reports, technical specifications, and similar past cases, all before they could even begin designing a solution.

Enter AI!

It quickly became clear that modern Large Language Models could significantly accelerate this initial research and analysis phase, so that engineers could focus their expertise on designing and implementing solutions – the work that truly requires human creativity and judgment.

However, the challenge was big. Root cause analysis for complex automotive issues involves:

• Reviewing hundreds of pages of technical documentation

• Searching for similar issues in private database

• Asking the "oldest" colleagues (or yourself) if something similar happened already

• Analyzing test data from multiple environments

• Connecting seemingly unrelated factors into coherent failure patterns

All of this traditionally required an engineer's trained eye and experience (precisely the resource in shortest supply.)

Impossible for AI to automate this process right away.

But then, something became clear: what if AI could work through the night, doing the heavy lifting of research and initial analysis, so engineers could start their day with a structured understanding of the problem?

And that’s when the solution unfolded!

The Solution: A 24/7 Problem Engineering Assistant

Now that the AI didn't have to come up with a result right away, but could literally spend hours on a problem, Valentin rebuilt the process from scratch, mimicking what engineers actually do. For months, he had been building complex AI approaches that didn't match real-world engineering workflows. The results were unstable and hard for engineers to trust.

Then it hit him: he needed to strip everything down and rebuild by mimicking the exact engineering process. He deleted everything and rebuilt a simpler version in just 48 hours.

Sometimes you need to remove complexity to unlock value.

Breakthrough #1: Mirroring the Engineering Mind

Valentin's approach focused on creating a workflow that precisely mirrored how their best engineers analyze problems:

Step 1: Problem Input Simple technical problem description – the system that failed, how it failed, and when/where it failed. Nothing fancy, just what engineers would normally write in slide decks.

Step 2: Three-Stage Analysis Pipeline

• Deep Research: Parallel web searches to find potential causes

• Main Cause Extraction: Converting unstructured information into a structured format

• Subcause Identification: Digging deeper into the root issues for each main cause

The Magic Ingredient: Verifiable Sources For every cause and subcause, the system provides clickable URL sources. No black box answers – every insight is grounded in actual documentation.

The Output: Visual Understanding A structured root cause analysis report with a beautiful fault tree diagram. Engineers can click any node to see the original sources.

To see the system in action, check out the video below:

A senior engineer's verdict? "This is exactly how I would research this myself – but it would have taken hours instead of minutes."

Sure, engineers do not rely only on web search. They mostly use their own experience and private data, but it turns out LLMs are pretty good deep web searchers. Plus, the the web is full of problems that are similar to the one we have.

Connecting it to private data would be the very next logical (and more complex) step for this project!

Breakthrough #2: Running at Scale While You Sleep

The baseline AI agent was impressive, but Valentin faced a new challenge: scale. With hundreds of issues in their backlog, running analyses one by one wasn't practical.

His solution? Transform the tool from a one-off solution into a continuous research engine that works overnight.

The Automated Pipeline:

• Connects to a Google Sheets issues backlog

• Pulls new technical issues automatically

• Processes them one by one through the analysis engine

• Updates the sheet with links to completed reports

• Runs until the backlog is empty or hits a daily limit

Smart Resource Management:

• Langgraph as the main orchestrator

• OpenAI's models for the heavy analytical work

• Gemini Flash for lighter tasks like formatting

• Pydantic AI for structured data extraction

Engineers could now start their mornings with fully researched problems instead of raw issues – like having a tireless assistant working through the night.

Results

The real-world impact of this system goes beyond just processing more issues. While it doesn't directly solve problems (that still requires human expertise), it eliminates the research bottleneck that previously consumed a significant portion of engineering time on complex issues. It's like having a junior engineer work overnight to prep everything for the team.

• Time Reclaimed: Engineers can now spend their mornings reviewing structured analysis instead of starting from scratch

• Shrinking Backlogs: Issues that would have sat untouched for weeks now arrive on engineers' desks fully researched

• Faster Solutions: Teams get to solution design quicker, skipping hours of preliminary research

• Continuous Processing: Set it up once, and it keeps working night after night

The most significant change is in both the efficiency and quality of engineering discussions. With the fault tree diagrams as a starting point, teams can spend less time debating possible causes and more time evaluating evidence and designing solutions.

It's also an excellent training tool. New engineers can see how complex problems break down into interconnected causes, learning troubleshooting methodology through real examples.

While still early stage, this demonstrates what is perhaps the true promise of AI in engineering: not replacing engineers, but removing the tedious parts of their jobs so they can focus on what humans do best – creative problem-solving and innovation. It's AI augmentation at its best.

6 Hard-Earned Lessons for Non-Technical Leaders

After two years of nighttime AI experimentation, here are the key principles that made this possible – even without coding expertise:

1️⃣ "Mirror real business processes"
Don't invent new AI-centric workflows. Digitize your existing methodology. When the system mirrors how your best people already work, adoption becomes natural.

2️⃣ "Stick to proven patterns"
Focus on established AI approaches rather than experimental techniques. This delivers reliable results instead of impressive but unstable innovations.

3️⃣ "Sometimes you need to remove complexity to unlock value"
Don't be afraid to hit reset. Valentin's most productive development came after deleting months of work and rebuilding from scratch in just 48 hours.

4️⃣ "Focus on one persona and one pain point first"
Solve exclusively for one specific user facing one specific problem. Deliver exceptional value in one area before expanding.

5️⃣ "Visualization builds trust"
Engineers need to verify before they'll trust. Make the AI's reasoning transparent and results inspectable.

6️⃣ "Start small, iterate quickly"
Begin with a minimum viable approach and incorporate user feedback in each iteration. The AI space is evolving rapidly, so quick iteration is essential.

Conclusion

You don't need to be a coding expert to create meaningful AI solutions.

Focus on real problems, understand actual workflows, and embrace simplicity – that's how even non-technical leaders can drive significant improvements with AI. Whether you're in engineering, marketing, finance, or any other field, these principles remain the same.

AI is most powerful when it's solving actual pain points that matter to your business – not chasing the latest hype cycle.

Keep innovating and…

See you next Friday!
Valentin & Tobias