Projeler

An autonomous pipeline that turns a single video idea into a finished 9:16 vertical short: LLM scenario breakdown, identity-locked per-scene frames, animation, audio, assembly and upload. It runs on a single 16 GB GPU, and that constraint shaped most of the design.

Automated VAT reconciliation against ledger records, on a backend built with CQRS and Domain-Driven Design. I built an LLM agent with LangChain and RAG that extracts data from invoices, detects inconsistencies and suggests corrections.

Backend handling 10K+ requests at sub-150ms latency, with data integrity ensured by 150+ unit tests and 60%+ code coverage. The technical articles on clean code and the anemic domain model came out of this project.

LLM agent built on NestJS.

Backend with JWT authentication and Entity Framework Core.

Built the backend of an emergency-response chatbot in a 4-person team; integrated the OpenAI API for the conversational layer, working straight from the official docs rather than an existing example. Deployed on AWS for stability under high load. Coordinated via Trello with weekly sprints; completed 45+ tasks on time.

Built as a real CRUD app on Supabase (Postgres + Auth + RLS), not a static site, with GitHub Actions CI and conventional commits. The stack was chosen deliberately: Next.js, Supabase and Vercel were tools I had no production experience with, and learning them on a real piece of work seemed the most honest way to close that gap.