At SMILINNO - PersianPoetry, I built a Neo4j knowledge graph from source data through schema design, crawling, and streaming ETL, scaling it to ~5M nodes and ~16M edges. I also built nine Airflow DAGs for data pipelines, graph repair, vectorization, and validation.
I designed a LangGraph/Langfuse query-planning layer and built the FastAPI backend for a Persian-poetry chatbot and interactive game. After replacing an unstable ReAct orchestrator with intent and entity routing, I helped bring intent latency from 60 seconds to 2 seconds and meaning retrieval from ~50 seconds to ~22 seconds.
I also built a supervised fine-tuning dataset factory and evaluated 17 Qwen2.5-7B configurations, improving masked perplexity from 25.02 to 21.64. Earlier, I delivered web and mobile applications at Metafan and worked on data and simulation systems, including graph curation workflows.
