At Resanai Company, I built a fully local, real-time speech-to-speech assistant from scratch, integrating streaming ASR, LLM inference, TTS, VAD, and browser audio over WebSockets. I implemented low-latency streaming and interruption handling, reaching 916 ms p50 end-to-end latency.
I also developed an LLM assistant with tool calling and a RAG pipeline, and built a Persian university regulations retrieval system using hybrid dense and sparse search. It achieved Recall@10 of 1.000 on a 38-query benchmark; I’m currently extending it with grounded answer generation.

