At Amnafzar Gostar sharif, I built a LangGraph pipeline that turns Persian questions into Splunk queries for MCI’s SOC team. Dynamic retrieval improved reference-match accuracy by 23 points over the no-RAG baseline on a curated 104-query set.
I also replaced static few-shot examples with retrieval from pgvector, improving query-generation accuracy by 12% over the static baseline. A bounded reflection loop repaired failing queries, while a dual-validation framework evaluated generated queries with an LLM judge and live execution.
For MCI’s internal teams, I built a multi-agent chatbot that provides natural-language access to PostgreSQL, MongoDB, ChromaDB, and MinIO. I also developed a document-intelligence pipeline that extracts structured content from financial documents and contracts.
At Agentco.ai, I worked on prompt optimization and medical RAG pipelines. At Samen Research Institute, I developed a DistilBERT and temporal-CNN approach for fake-news classification and automated data collection pipelines for NLP model training.
