How I got here

I did not set out to become an AI engineer. The path started with a question that had nothing to do with code: how does the human mind work? Growing up in Ahmedabad, I was the sort of person who took things apart to understand them. Radios, old phones, anything with a mechanism. The brain was the one machine I could not open.

My undergraduate degree in computer science at Ahmedabad University was where the two interests met. Programming gave me tools for modelling complex systems. Cognitive science gave me the most complex system of all. My first real project combined them: a simple neural network trained to classify emotional states from physiological signals. It was crude, but it worked well enough to commit.

I moved to Birmingham deliberately. The university has strong groups in both AI and cognitive science. Professor Christopher Baber’s work on human factors sat exactly where I wanted to be. My thesis, GraphMinds, grew from a conviction that AI systems should show their working the way a good researcher does.

EEGSpeech brought me closer to clinical work. When you build a system designed to help people with locked-in syndrome communicate, every percentage point of accuracy matters differently. It is not a metric on a leaderboard. It is the difference between someone expressing a thought and not being able to.

Alongside the academic work, I spent time at Inzeitech deploying models that served over 150,000 users. Production taught lessons that research alone does not. Latency matters. Edge cases multiply. The gap between a model that works in a notebook and a system that works in the real world is vast.

The safety work came from a growing awareness that the systems I was building had become genuinely capable. Jailbreak-Eval, my adversarial testing framework, grew from a practical need: if you deploy these models, you should know where they break.

I keep returning to problems where structured reasoning meets unstructured data. Knowledge graphs appear in nearly everything I build because they sit at that boundary. They impose structure on messy information, make reasoning traceable, and pair well with language models that handle the ambiguity rigid schemas cannot.

London suits the work. The city has a dense concentration of AI labs, research groups, and startups. I spend most of my time building, writing, and reading papers. Winning the Epiminds multi-agent hackathon was gratifying less for the prize and more for the confirmation that the ideas I had been developing translated well under pressure.