Senior NLP Engineer & Machine Learning Engineer

Built an agentic AI layer for a support assistant using a ReAct loop in LangGraph, a tool registry with argument contracts, and GPT-4o function calling. Implemented deterministic guardrails for reversible operations and confirmations, achieving tool selection accuracy of 0.91, argument accuracy of 0.87, stopping accuracy of 0.94, and increasing resolution rate from 49% to 58% with zero irreversible errors.

Designed the retrieval core for a multi-tenant RAG support system using BM25 in Elasticsearch, bge-m3 embeddings in Qdrant, and RRF fusion with tenant isolation. Fine-tuned a bge-reranker-large model on 180K log-derived pairs and added DeBERTa-v3 grounding, increasing faithfulness to 0.91, Recall@10 from 0.68 to 0.90, and resolution from 31% to 49%. Also created a knowledge-base audit pipeline using UMAP, HDBSCAN, bge-m3, and NLI to identify unresolved-dialogue patterns and article contradictions.

Developed production NLP systems for fraud prevention, search, listing moderation, and multilingual attribute extraction. Built a marketplace-chat scam classifier using XLM-RoBERTa, deobfuscation, and a character-level CNN, improving precision to 0.87 and recall to 0.71 while reducing false positives from 8% to 1.2%; fraud-related support tickets decreased by 38%.

Rebuilt search with Elasticsearch BM25, FAISS ANN retrieval, RRF fusion, a contrastively fine-tuned LaBSE bi-encoder, and an XLM-RoBERTa reranker deployed through ONNX INT8. Also automated moderation across 23 categories with XLM-RoBERTa and CatBoost, retrained via Airflow, and delivered a nine-language NER and brand-linking pipeline that improved listing completeness and conversion.

Master’s degree in Computer Science and Computer Engineering.