A senior engineer who treats reliability as a feature.

I started writing HTML by hand in Prishtina before frameworks existed, on machines where a reload cost you ten seconds of your life. That is probably why performance instincts stuck: I learned early that speed is something you feel, not something you measure in a report nobody reads.

Nineteen years later the tools are unrecognisable — TypeScript, Next.js, Prisma, vector search, streaming LLM responses — but the job is the same. Somebody has a workflow that hurts. My work is to understand it well enough to build software that removes the hurt without creating three new ones.

Most of my career has been spent in the unglamorous middle: taking a half-finished system, working out why it is slow or fragile, shipping the missing 30%, and leaving behind documentation that means I am not needed next quarter.

The AI work arrived the same way — not as a research interest but as a delivery problem. Teams wanted document intelligence, voice handling, and proposal automation that behaved predictably under real volume. That means structured outputs, evaluation runs, cost telemetry, and clear escalation paths. The model is the easy part.