Vladislav Khomiakov

I’m an AI Engineer with 5+ years of commercial experience building end-to-end AI solutions and driving real business impact. I own the full development lifecycle—from turning business goals into ML tasks, to training and fine-tuning models (LLM, NLP), designing agentic architectures, and deploying scalable high-load services.

At CrowdStrike, I extended an AI agent integration with Qualys VMDR to cover infrastructure assets, vulnerabilities, and configuration management. I replaced RAG-based QQL generation with deterministic tool selection, driving ER “from 50% to 1%,” and I deployed a dedicated MCP server using Docker, Kubernetes, and Helm. I also redesigned the agent loop into a Multi-Agent ReAct workflow with LangGraph, improving Faithfulness “+80%” and ER “50% to 1%,” while migrating the backend from Django to FastAPI for better scalability.

Previously at Langdock, I led NLP and RAG work for an analytics department and production meeting workflows. I built a document ingestion pipeline with Qdrant and hierarchical chunking (improving Precision@5 by +0.21 and MRR@10 by +0.24), fine-tuned BGE-M3 on an internal wiki, and implemented hybrid retrieval with dense/sparse representations and Reciprocal Rank Fusion (Precision@5 “0.89” and MRR@10 “0.91”). I also deployed Qwen3-30B with Docker and vLLM (prefix caching, FP8 KV-cache quantization, continuous batching) to raise throughput by 40%, and I reduced meeting summarization latency “from 20 minutes to 3 minutes” using an asynchronous hierarchical summarization pipeline and a production-grade FastAPI service.

Earlier, as an ML Research Engineer at DeepPavlov, I fine-tuned a pix2tex image-to-LaTeX transformer on 120k formulas, reducing symbol error rate from 9.4% to 3.5%. I also developed a hybrid scientific search engine using BM25 + SciBERT with OpenSearch and FAISS, improving MRR from 0.31 to 0.54. My approach is to rapidly dive into new tasks and drive research-driven innovation in production, backed by a strong fundamental foundation at MIPT.