KIPRIYAN PANAYOTOV

I'm 20, with 2 years of experience as a software engineer. I started programming at 10 at Telerik Kids Academy, where I spent three years on game development and algorithms. I went on to SPGE John Atanasov to become an engineer. AI arrived when I was in 11th grade, and in 12th grade I built my entire diploma project with the models of the time. They were far from smart, but I found ways to make them do the work.

The same month I finished high school I joined Encorp, first building AI automations, then moving to a team serving a large European payments institution — bank accounts, card issuing, merchant acquiring and international transfers. There I learned to build and connect APIs, deploy what I build and, above all, talk to clients. My work has always been about using AI in the most advanced way, so it does the work for me.

In May 2026 the first truly capable coding models arrived, and I spent all my time learning to get the most out of them. At first the bottleneck was reviewing what the model wrote. Soon the output got so good that it moved to gathering the right context for the model. The question was no longer how to build something — it was what to build.

By September my whole team was delivering its work with AI agents, and I went looking for the next bottleneck. It was our project manager, buried in the client's endless questions. I built a platform of AI agents that answered them from everything I could gather — client knowledge, codebases, tasks and git history: what runs where, how far along each task is, how loaded each developer is. I added document generators for change requests and process overviews for third-party auditors, taught to present things the way he talks to the client in emails, meetings and chats.

Then the bottleneck moved again. The client started writing tickets with AI: vague, full of terms we had never heard, with acceptance criteria that didn't understand the process they were changing. Their agents started leading the conversation too, so every email came with a hundred questions, since their agents didn't have any context. A chat tool was no longer enough — AI had to understand each project, work out what the client actually wanted and answer in our tone, automatically.

I pulled everything we had on the client — documents, meeting notes, code, emails, chats and tickets — into one project knowledge base. It was too long for any person to read, so I built a test set of real tickets and questions it had to help with, ran it again and again, and kept only the information that earned its place. Then I connected it, along with all our other sources, to our agents through a server my colleagues use, and set up daily jobs that keep it current with new conversations, tickets and shipped features.

Now our project manager answers twenty questions at once and only checks the answers before sending. Every developer can see what a ticket really asks for and what the logs are showing; code review is automated, and staging deploys and testing mostly come down to asking an agent. The bottleneck is back where it belongs — human review.