CoreStory

CoreStory helps companies understand and modernize legacy codebases that nobody fully understands anymore. It generates artifacts and insights for developers and architects.

These codebases are too large and complex for generic AI tools or manual analysis alone, so CoreStory was built to handle what neither could.

The product came out of a major pivot at Crowdbotics and became the company's entire focus.

Enterprise clients had massive systems built decades ago by people who no longer work there. Written in legacy languages like COBOL, FORTRAN, and early C++, these codebases have become black boxes. Modernization projects stalled because nobody could answer: "What does this system actually do?"

Companies were spending €50,000–80,000 rehiring retired developers just to interpret their own code. Delaying projects by months.

Why companies urgently need to modernize:

At the start of this project I became the sole designer at the company. Other teams were waiting on designs and direction, so I had to find ways to move through the research and discovery phase fast without cutting corners. I used AI to transcribe and synthesize interview findings in real-time, spending just 10 minutes per call to extract structured insights while maintaining accuracy and depth.

Users wanted to change the prompt for each artifact, so I needed to understand how each one was actually created. I put together a diagram to map out the full AI workflow. Looking back, the fact that customers were asking to tweak prompts was an early sign of a bigger problem. They didn't really trust the AI output. We'd dig into that in Phase 3.

The timeline was tight, so I maintained a rapid design cadence, delivering work for review three to four times a week. Deliverables ranged from small refinements to entirely new features, and when we needed fast validation, I turned key designs into clickable prototypes and tested them internally with teams like sales.

We discovered the trust gap late. In early testing, developers praised how everything looked but hesitated to actually use the outputs. Further questioning revealed the cause: bad experiences with other AI tools and constant news about hallucination had created deep skepticism. If customers didn't trust what the platform produced, the project was dead in the water. We had to solve this fast.

Instead of presenting AI outputs as definitive, we displayed confidence scores on every artifact. The logic was counterintuitive: admitting uncertainty actually builds trust. Users could immediately see where the AI was confident and where it wasn't, so they knew exactly where to focus their review.

If users didn't trust the output, they needed the power to change it. We let them provide context upfront, adjust the prompt, and edit parts of artifacts directly. Guardrails kept them from breaking the underlying analysis, but the key shift was psychological: users went from receiving AI output to steering it.

Every AI-generated insight links back to the exact source code that informed it. Users can click through and verify any conclusion themselves. The message was simple: don't take our word for it. This was the single most effective trust builder in testing.

Even if users trusted the AI, they needed to know nothing was permanent. Full version history with rollback, comparison, and visual diffs meant that trying the platform was risk-free. If the AI got something wrong, users could always go back. This removed the biggest barrier to adoption: the fear of irreversible mistakes.

As the sole designer working at a rapid pace, I was focused on shipping polished work fast. Everything looked great, and the feedback reflected that. But looking great wasn't the problem. When early testers hesitated to actually use the outputs, it forced me to confront something I hadn't designed for: trust. That moment changed the direction of the project. It taught me that with AI products, what users see matters far less than whether they believe it.

On a personal note, I never expected my curiosity about large language models to become directly useful at work. I had been experimenting with LLMs and prompt engineering on my own time, and when this project needed someone who could collaborate with the AI team on prompt workflows, that personal interest suddenly had professional value. It's a reminder that following what genuinely interests you tends to pay off in ways you can't predict.

I built the design system from scratch. I worked closely with the frontend developers to make sure components in Storybook matched Figma exactly, same naming, same variations, so there was never any confusion between what I designed and what they built.

The sudden pivot was difficult for the team. Everyone had to abandon projects they'd invested over a year in. To help lift spirits, I created this video in about four hours to celebrate our new direction. I also had one-on-one conversations with developers who felt their work had been discarded, helping them see how their efforts still mattered and keeping morale high during a challenging transition.