Dmitrii Konyrev

I build LLM systems that companies can actually trust in production: agents, RAG, and the evaluation that proves they work.

Most AI projects fail not because the model is bad, but because nobody can tell whether the output is right. I start with an evaluation harness, then build. Quality is measurable from day one, and you can show a stakeholder why the system can be trusted instead of just asserting it.

What I can help you with:

- LLM agents and multi-agent systems (LangGraph, Atomic Agents, OpenAI and Anthropic models), with tool use, memory, routing, and human approval gates.

- RAG and semantic search: embedding pipelines, chunking, hybrid retrieval, reranking, on pgvector, Qdrant, or OpenSearch. I am the author of EmbeddingStudio, an open-source framework for turning vector databases into production RAG.

- Evaluation, guardrails, and AI risk: LLM-as-a-judge, DeepEval and DeepTeam, hallucination and grounding checks, red-teaming and prompt-abuse scenario packs.

- Inference cost optimization: model choice, prompts, and routing tuned to cut cost 2 to 5 times while holding quality.

- MLOps: Python, FastAPI, PyTorch, Docker, MLflow, AWS and GCP.

Background: about 10 years in ML and AI, 9 of them leading teams. Currently CTO of Argmin AI. Previously ML Team Manager at SuperAnnotate, Risk Modeling Team Lead at Bank DOM.RF, and R&D Team Lead at Sberbank, where I led 6 data scientists and 5 engineers and shipped an NLP service that cut corporate loan decisioning from 7 days to 7 minutes.

Typical deliverables: a deployed agent or RAG pipeline with documentation, an evaluation suite with real quality gates, an architecture review, or a technical due diligence on an existing AI system.