zored132

Developed a chat and call center request classification system by fine-tuning ruBERT in PyTorch. Managed training and experiments in Weights & Biases, versioned models with DVC, and exported the model to ONNX. Achieved an F1 score of 0.86 for request routing and reduced the workload of first-line support operators by 25%. Designed a multi-agent system for processing small-business loan applications. Implemented orchestration with LangGraph using state graphs and conditional routing between agents. Used Qwen2.5-72B-Instruct as a self-hosted LLM served via vLLM, along with function calling and few-shot prompting. Stored agent states in Redis and logged execution traces in LangSmith. Reduced the initial application processing cycle from three days to six hours and lowered scoring-related operating costs by 35%. Built a RAG system for an accounting assistant using Qdrant as the vector database. Implemented hybrid search with BM25 sparse retrieval and cross-encoder reranking, and orchestrated the pipeline with LangChain. Achieved Recall@5 of 0.87 and reduced the response time for complex queries from four hours to 15 minutes. Implemented an automated LLM response quality evaluation framework for a seller assistant within the Tochka ecosystem. Used DeepEval and RAGAS, custom LLM-as-a-judge metrics powered by the OpenAI API, a golden dataset, and regression gating in GitHub Actions. Reduced the share of incorrect production responses from 12% to 4%. Tech stack: Python, PyTorch, Transformers, PEFT, LoRA, Accelerate, vLLM, Docker, Kafka, FastAPI, ONNX, MLflow, OpenAI API, Kubernetes.

Built an NER pipeline for extracting entities from legal documents by fine-tuning ruBERT and optimizing hyperparameters with Optuna. Improved F1 score from 74% to 96%, reduced manual processing by 40%, and saved legal teams approximately 15 hours per week. Optimized an ETL pipeline by migrating orchestration to Apache Airflow, implementing asynchronous I/O with asyncio and aiohttp, eliminating PostgreSQL bottlenecks through indexing, partitioning, and query optimization, and adding Redis caching. Increased processing speed by 1.8x and reduced analysts’ time-to-insight by 50%. Developed a RAG-based legal document search and answer generation system from scratch for a corpus of more than 500,000 documents. Used Qdrant as the vector database with hybrid dense and sparse retrieval and orchestrated the pipeline with LangChain. Improved MRR from 0.62 to 0.90. Deployed LLMs with vLLM and a ruT5-based ranking model exported to ONNX. Containerized both models with Docker and integrated them into production through FastAPI. Tech stack: Python, PyTorch, Transformers, BERT, LoRA, vLLM, ONNX, Docker, FastAPI.