ML Engineer — LLM systems, agents, and retrieval. I care about the parts that have to run after the demo is over.
ML Engineer at Sber, building multi-agent systems for codebase analysis and code standardization. Before that — computer vision (Eco-Scan) and LLM routing (Exine). Competitive-programming background — prize at «Я-профессионал», and I still reach for C++ when a millisecond is on the line.
- Agent systems that survive real codebases. LangGraph state machines, tool orchestration, recovery paths — not happy-path notebooks.
- Retrieval that retrieves. Hybrid BM25 + dense (BGE-M3) + cross-encoder rerank, fused with RRF. Cosine over a single embedding is where the problem starts, not where it ends.
- Memory for long-horizon agents. STM/LTM hierarchy, MemGPT-style paging, causal-violation tracking. (I gave a talk on Letta / MemGPT internals.)
- Latency. Profile first, cut lines second. C++ on the hot path, Python for everything else.
Celerity · C++ Compact, training-free multimodal search. Built to be small and fast, not to wrap a framework.
CNN · Python Real-time sign-language translation — CV running inside the UI, not on a server somewhere.
Secure_chatbot · Python Guardrailed LLM chat — the boring, necessary half of shipping an assistant.
Python · PyTorch · LangChain / LangGraph · scikit-learn · C++ · FastAPI · Docker · PostgreSQL · Redis · Linux
Reasoning via RL — the DeepSeek-R1 lineage, GRPO and what comes after it. Agent memory past plain RAG.
orvune.tech · Telegram · kirillzinchenko2006@gmail.com