When I started thinking about my thesis I only had a rough sense of the area I wanted to explore. Something about AI systems and how they behave in the real world something about agents and workflows something about reliability. Nothing felt settled. It took a few twists to find the topic that now feels like the right fit.
The first idea that didn’t stick
My earliest plan circled around multi agent systems and automated workflows. It sounded exciting but each conversation made me realise how undefined the space still was and how hard it would be to turn into a focused six month project. The topic kept drifting and the setup around it became less certain. In hindsight it was the right call not to commit. Letting it go opened the space for something clearer.
Finding a direction through testing
The turn came when I started looking at the proposed topics around testing AI systems. At first it felt almost too narrow but the more I explored the literature the more it expanded into a full landscape of ideas. I spent weeks reading about model behaviour reproducibility drift and all the ways AI models can shift in practice. During that time the topic kept reshaping itself. Some days it felt promising and other days it felt frustrating because every angle either seemed too broad or too shallow. But slowly a pattern formed.
A necessary adjustment
One piece of feedback that changed the course early on was the reminder that models often don’t evolve as dramatically as we assume. Treating them as constantly shifting targets would have made the whole thesis stand on shaky assumptions. Broadening the scope to include not only models but the systems around them opened a much better path. It grounded the project and made it more realistic.
Conversations that shaped the project
Regular meetings helped turn loose thoughts into something structured. Every session pushed the idea a bit further clarified what mattered and cut away what didn’t. It was steady and constructive and it made me more confident that this was becoming an actual research direction and not just a cluster of notes.
The moment it clicked
The real turning point was building the first small MVP. Seeing a minimal version of the framework run end to end made the whole thing tangible. It showed that the idea is workable testable and worth developing. That moment shifted the project from speculation to something I can genuinely build over the next months.
Where it stands now
The thesis has become a study of behavioural verification for AI systems. How to define expected behaviour how to rerun models under controlled variation and how to record the hidden decisions that explain why behaviour changes. It grew out of several dead ends a lot of reading and many adjustments but now it feels coherent and grounded.
It took a while to get here but the process shaped the topic in a way I wouldn’t have managed alone. And with a clearer direction the next months already feel more focused than the first ones ever did.
