Keeping the AI Train on the Tracks: Lessons on Intentional Innovation with Intelligent Tools 

Our team explores how generative AI technologies will transform design and manufacturing customers’ messy and complex challenges into efficient and transparent workflows. Until recently, some of our complex ideas were held at a standstill, faced with roadblocks like skill or knowledge gaps. It’s not uncommon for this to be a regular occurrence forcing us to move at a slower pace and possibly run out of steam entirely.

This way of working has been completely disrupted using resources like Cursor and other AI coding tools. As a group of Mechanical Engineers with limited software engineering experience, so much is now possible with AI. We’ve never been short on ideas, but the skills and resources to realize them is where we fell short before using AI.  Here’s four perspectives on our experience working in an entirely new reality!

Part 1: When Research is too Abstract: Can Vibe Coding Help?

Brandon Cramer

In our research, we’re exploring the many ways AI, particularly LLMs, might augment a design engineer’s workflow. While there are distinct phases in any design process, the overall journey is far from linear, which means there’s enormous potential to implement AI in interesting ways.

As a team, we usually set aside dedicated time to brainstorm and share ideas. But over the 4th of July holiday, I found myself highly motivated to try something new on my own. During some downtime, I opened Cursor, the vibe-coding tool that had recently been made available to Autodesk employees. I already knew what it was like to interact with ChatGPT, Claude, Perplexity, and others, but testing Cursor for the first time completely caught me off guard.

I found myself exploring ideas in an enticing new way, and I was dying to tell someone about the flood of possibilities that kept coming to mind for our team. Just like all the blogs, articles, and tech influencers have been saying, there’s a new frontier in software development happening.

But as a people manager, my immediate focus was on how this would impact our way of researching rather than what we would research. This scrappy group of mechanical engineers has an opportunity to innovate faster than ever before, and it left me wondering: With this new capability, how might we generate and realize our ideas in a more collaborative way that positively impacts our broader team? With these questions swirling around in my head, I turned to Dani, our go-to-applied AI expert, to devise a plan to get the whole team onboard a new way of working.

Dani Grandi

In research it’s never as easy as ‘diving into a project headfirst,’ as much as everyone would like to believe so. Usually, the organization frames the problem in very abstract terms, the team defines it in somewhat general terms, the sub-team tries to narrow it down, and the issue you’re assigned still feels too abstract once you start tackling. Matt Might, in “The illustrated guide to a Ph.D.,” cleverly depicts how a Ph.D. candidate will have to narrow down the scope and hone in on a very specific problem, but the challenge becomes keeping an eye on the big picture. Research projects suffer from the same zoom-in/zoom-out problem: the original problem to be solved is always too broad, and the final output of the research is always too narrow. Coupled with the constant reorgs, and the unavailability of engineering resources or data, this leads to a lot of disparate projects that are often too far from each other to build a whole picture.

“The illustrated guide to a Ph.D.” can also be used to depict research projects, and the connection between individual contributions and high-level organizational aspirations.

Can vibe coding help? Somewhat.

Primarily, tools like Cursor have allowed non-software engineers to emulate software engineers when they are booked onto other projects. This independence from software developers has allowed our team to create very quick prototypes, often not much more than clickable mockups, but functional tools that allow us to quickly provide an initial working technical implementation of some new technology to solve specific user needs. Cursor takes this process down from six months to two weeks. And the best part: if the experiment does not pan out, or if the team’s goals change, throwing out the code is completely painless, since it was never about the code, but about quickly testing a functional prototype.

While development speed is likely the largest benefit of vibe coding, another important aspect comes yet again as another independence, this time as a lack of reliance on the UX team. Being able to leverage MCP clients to call research prototypes (hooked up as MCP servers), allows our team to defocus on the UI, and instead focus on testing the functionality of our research prototypes. This means, once again, that we are less reliant on other teams and their busy schedules, to deliver working prototypes.

MCP also comes with some other benefits: it’s become easier to share code amongst researchers, as this common protocol can be easily adopted. This allows researchers to still work in silos and develop independent prototypes that solve completely different problems with completely different tech stacks and technologies, but eventually bring them together into a fully working workflow to demonstrate a big vision more easily.

The elephant in the room: code quality. By far, the biggest challenge we faced was the downstream integration of these vibe-coded prototypes. When Cursor writes out 10 files with 500 lines of code each, debugging becomes very time consuming, and this is the train that needs to be managed.

Part 2: Crossing the Skills Gaps with AI as a Thought Partner

Kevin Acker

About a year ago, our project team identified a knowledge gap in Model-Based Systems Engineering (MBSE) and I was given the chance to begin building subject matter expertise to help fill it. After working through online courses, I reached a point this spring where MBSE’s principles and methods felt familiar enough to start applying them in real projects.

Then a new challenge emerged: we needed to build a set of prototype tools to demonstrate a critical workflow. My task was to automate the trade space analysis process with an AI-powered solution. Understanding MBSE in theory was one thing; turning that theory into a working prototype was another and it demanded programming experience I didn’t yet have.

A few years ago, that might have stopped me in my tracks. But today, with the rapid rise of AI tools I turned to Cursor and other LLMs not as simple answer generators, but as collaborative partners that could help me explore, reason, and build the code I needed.

Cursor helped me bridge two gaps at once – deepening my understanding of MBSE while translating that understanding into working code. Instead of poring over documentation or stumbling through syntax, I could ask questions, test ideas, and see them translated into functional snippets in real time. Cursor even validated and refined the code, transforming what could have been weeks of trial and error into a rapid cycle of learning and building.

The results were immediate. I created a prototype for an LLM-powered trade space analysis solution, something that would have been beyond my reach prior to the release of these latest AI tools. More importantly the process reshaped how I think about my role. I realized I didn’t need to be an expert programmer or a seasoned MBSE practitioner to make progress. Expertise is no longer a prerequisite for innovation. By combining curiosity, a willingness to learn, and AI as a collaborator I could venture into new territory and build something tangible at the same time.

That experience taught me a larger lesson; AI doesn’t just accelerate the tasks we already know how to do, it allows us to cross into entirely new disciplines. When engineers use AI to step into unfamiliar domains, the traditional boundaries between specialties begin to dissolve. The result is a future where teams innovate faster, collaborate more effectively, and bring ideas to life that might otherwise have stayed parked on the station.

If AI can turn one person’s learning curve into a working prototype, Tyson’s teardown example shows the exponential impact when entire teams apply the same mindset to whole areas of missing expertise.

Tyson Fogel

As innovators and researchers, we always seek to understand, we ask questions, and we mine for context to better grapple with the problems at hand. We prototype things to test hypotheses, we pause on sticking points to seek answers, and we apply empirical evidence to move forward.

But we are often left with more questions than answers…

This experience is not unique to any single industry or profession. It is for this reason that we often find ourselves working on cross-disciplinary teams, taking advantage of the learned experience and diversity of perspective of others.

What happens when we don’t have the right experts in the room, or are missing context/additional information? How do we overcome these situational knowledge gaps? In the past, our reasoning would be delayed as we sought to connect with the right person and dive deeper into a topic to find the answer, but abstractions are made, and we can never have enough context. I can confidently say, this new wave of AI tools upends this dynamic.

Over the last few years, I have conducted a series of physical product teardowns. These reverse engineering exercises were undertaken to better understand how Autodesk could better serve the design and make industries through their software. I wanted to learn more about how assemblies were built from the ground up, the trade-offs being made by designers, and how these unearth the tacit knowledge that drove design decision-making. The early teardowns were tricky, and we always wished we had the designers and product experts in the room as our observations were anecdotal at best. During our most recent teardown activity, however, we developed a series of custom GPT personas primed on the disciplines, experience, and knowledge needed to design the product of interest. We queried this team of ‘experts’ to better qualify the observations being made during our work and backfill the context we needed.

The power of this is not to be understated as it allowed us to gain information rapidly and have a higher level of confidence in the data points we were gathering. Further, we were able to leverage these learnings and the data gathered from these research activities to accelerate the suite of prototype agentic framework we were already creating. The GPT personas gave us a benchmark, a reference and a contextual resource that is still actively being used today in our research, despite it being months later.

In this example, our team acted as the mediator between the physical world and digital team of experts. Although the personas offered direct and impactful insight, they were limited by the context and prompts we provided. It truly makes you wonder: what is possible if the same team of AI experts didn’t need a mediator, what next level of insight could they provide, and what uncharted territory could be discovered?

Conclusion: With Great Power…

As our team’s journey shows, the rise of generative AI has become the driving force of progress, much like a powerful locomotive forging ahead on new tracks. Where we once faced barriers of expertise, time, or resources, AI has enabled us to move forward with momentum, transforming initial sparks of inspiration into fully realized prototypes. Instead of waiting for the right skills, we now build them as we go, fueling a revolution in how ideas travel from concept to reality. AI empowers us to bridge gaps in expertise, collaborate across diverse teams, and accelerate innovation at a pace never before seen.

Yet just as a train’s speed demands careful , this newfound power comes with responsibility. The same tools that help us “emulate” new roles and propel innovation can also introduce risks: overconfidence in fast solutions, issues with code quality, or the temptation to bypass foundational learning. It’s easy to be carried away by the thrill of the ride, but meaningful progress requires balancing velocity with diligence, and experimentation with thoughtful review.

As we lay new tracks and push the boundaries of possibility, let’s remember: AI is not a shortcut, but a catalyst. The journey depends on us; collaborating, learning, and building with intention. If we do, we will not only transform how we work; it will redefine the destinations we can reach together.