At AIONOS, I self-hosted Qwen 2.5 7B inference on Azure and reduced time-to-first-token from over 11 seconds to under 2 seconds under a 1,500 RPM production load. I also reworked sequential LLM calls into parallel async execution, reducing end-to-end API latency by about 80%.
I architected a LangGraph fare-rule extraction system with hybrid GraphRAG retrieval across Qdrant and Neo4j. At GlobalLogic, I optimized live Twilio voice AI pipelines to reduce end-to-end latency by 30–40%; in personal projects, I’ve built conversational AI and agentic code-generation systems.
