ML / NLP Engineer

Develop and deploy an agentic support assistant that combines query routing, hybrid RAG, CRM and service-desk tool calling, and human approval for high-impact operations. The system automates approximately 35% of standard support requests.

Built hybrid retrieval over documentation and tickets using dense bi-encoder search, BM25, reciprocal rank fusion, cross-encoder reranking, ACL metadata filtering, and answer citations. Implemented controlled stateful workflows with Pydantic validation, RBAC checks, idempotency, retries, fallbacks, and audit trails.

Established offline evaluation using golden datasets, Recall@10, MRR@10, groundedness, answer relevance, escalation rate, and tool correctness. Delivered Dockerized services with CI/CD, health checks, and Prometheus/Grafana monitoring for latency, errors, token usage, caching, and retrieval quality.

Built an end-to-end RAG assistant providing natural-language search and question answering for engineers and support teams across technical documentation, procedures, incident reports, and resolved cases. The production service handled approximately 12,000 requests per month.

Created reproducible ingestion and indexing pipelines for PDF, DOCX, wiki, and ticket data, including structure-aware extraction, cleaning, semantic chunking, metadata enrichment, deduplication, and incremental updates. Fine-tuned a domain bi-encoder with mined hard negatives, improving Recall@10 from 0.68 to 0.86 and MRR@10 from 0.54 to 0.74.

Implemented hybrid dense and BM25 retrieval with RRF and cross-encoder reranking. Deployed FastAPI services using Qdrant, PostgreSQL, Redis, and Docker, reducing median retrieval and generation latency from 3.8 seconds to 1.7 seconds through caching, batching, and asynchronous processing.

Trained and deployed Transformer models for multi-class customer-request routing and named entity recognition from documents. Reduced manual routing of incoming requests by 42%.

Built reproducible training pipelines with data cleaning, temporal train/validation/test splits, class-imbalance handling, hyperparameter search, MLflow experiment tracking, and confusion-matrix error analysis. Improved Russian-language Transformer macro-F1 from 0.79 to 0.89 through mixed precision, dynamic padding, gradient accumulation, early stopping, and per-class threshold calibration.

Optimized CPU model inference by exporting models to ONNX, applying INT8 dynamic quantization, micro-batching, and FastAPI serving. Reduced p95 latency from 260 ms to 78 ms with less than one percentage point of quality loss, while adding Docker-based deployment, logging, Prometheus/Grafana monitoring, and drift checks.

Master of Science in Data Science with a specialization in Computational Neuroscience. Located in Almaty, Kazakhstan.

Bachelor of Science in Data Science. Located in Almaty, Kazakhstan.