ML Engineer
Neuro.net
Feb 2026 — PresentBuilt and deployed an agentic internal support assistant with evaluated hybrid RAG and workflow controls; automated approximately 35% of standard requests.
• Designed and deployed an agentic assistant for internal support teams, combining query classification, hybrid RAG, tool calling, and human approval for high-impact actions.
• Built retrieval over operational documentation and resolved tickets using vector search, BM25, RRF fusion, cross-encoder reranking, metadata-based access filtering, and source citations.
• Implemented stateful CRM and Service Desk workflows with structured outputs, audit trails, retries, fallbacks, idempotent operations, and controlled human handoff.
NLP Engineer
GRONICS
May 2024 — Dec 2025Delivered a production RAG assistant serving approximately 12K monthly requests, improving retrieval metrics and reducing median latency from about 3.8 s to 1.7 s.
• Built an end-to-end RAG assistant for engineers and support specialists across technical documentation, operating procedures, incident reports, and resolved cases.
• Developed ingestion and indexing pipelines for PDF, DOCX, wiki pages, and tickets: text extraction, cleaning, semantic chunking, metadata enrichment, deduplication, and incremental updates.
• Fine-tuned a bi-encoder on domain query–document pairs and hard negatives, improving Recall@10 from 0.68 to 0.86 and MRR@10 from 0.54 to 0.74 on an offline evaluation set.
ML Engineer
Fortech
Mar 2022 — Apr 2024Built and deployed text classification and NER models, raising macro-F1 to 0.89, cutting CPU p95 latency to 78 ms, and reducing manual request routing by 42%.
• Trained and deployed Transformer models for multi-class classification and named-entity recognition over customer requests and documents.
• Built reproducible data and training pipelines with cleaning, temporal train/validation/test splits, class balancing, hyperparameter search, experiment tracking, and error analysis.
• Fine-tuned Russian-language Transformers using mixed precision, dynamic padding, gradient accumulation, early stopping, and class-specific threshold calibration; improved macro-F1 from 0.79 to 0.89.
