I write backend and systems software, mostly where it meets LLMs. Recently: a KV-cache manager for LLM inference modelled on PagedAttention and RadixAttention, a limit order book with an allocation-free hot path, an HNSW vector index written from the paper, and a load tester that measures time to first token and inter-token latency on LLM endpoints. I also ship a language-learning app on the App Store whose speaking practice runs on speech recognition, LLMs and text-to-speech.
Fixes and Turkish localization merged in Apache Superset, MUI X, Element Plus and PrimeLocale. These days I'm contributing to cloud-init.
LLM inference, real-time audio, distributed systems, and getting software to speak Turkish properly.