Backend developer with 4 years in IT — Go, Python and C#, full-stack when the product needs it. MSc in Artificial Intelligence with honors (ITMO University, AI Talent Hub). I build ML systems and LLM agents — from an AI interview simulator running in Kubernetes to an operating system you talk to.
🏆 Winner of the τ²-Bench competition on AgentBeats (Berkeley AgentX) at the competition close.
Graduation thesis · team of 2 · Oct 2025 – Jul 2026 · live pilot at tensor-talks.ru
An AI interviewer runs the session, scores every answer and turns mistakes into a concrete preparation plan. 18 Go/Python microservices, three LangGraph agents (planner, interviewer, analyst), Kafka (Strimzi), Kubernetes with ArgoCD GitOps, Vault, Langfuse observability. The platform code is private; two public repos:
tensor_talks_demo · Jun 2026 Public showcase: architecture documentation, UI screenshots and GIFs, defense and meetup slide decks, the final thesis PDF.
tensor-talks-agentic-ai-course · Mar – Apr 2026 Agentic AI course capstone on the live platform, full source included: an LLM ingestion workflow for the knowledge base plus planner, interviewer and analyst agents (LangGraph ReAct) with guardrails, episodic memory and confidence-aware decisions.
Competition · Apr 2026 · agent-beats-tau2
🏆 First on the AgentBeats τ²-Bench leaderboard at the competition close (Berkeley AgentX) — 82.5% in the telecom domain. A purple customer-service agent for the airline, retail and telecom domains: a plain A2A service with no agent framework. The edge is distilled per-domain policy playbooks injected ahead of the task policy, with tool definitions parsed straight out of the system prompt.
Stack: Python, a2a-sdk, OpenAI-compatible LLMs, Docker, CI to GHCR.
Pet projects · Jul 2026 – now
Four apps that talk through Kafka (CloudEvents 1.0, transactional outbox). Each one is self-sufficient and exposes three facades: a REST API, events and an MCP server for LLM agents. The server side runs on a VPS; desktop clients connect over an SSH tunnel.
task-planner Tasks as recurrence rules — nine kinds including spaced repetition — materialized into daily occurrences. Desktop and Android are built from the same Go core and the same React UI, with two-way sync between devices. Go, Gin, GORM, Kafka, MCP, Wails, React/TS, gomobile.
pomodoro One-process pomodoro timer: a frameless always-on-top overlay and a day plan auto-built from task effort estimates. Go, Wails, React/TS, PostgreSQL, Kafka, MCP.
motivation-bot A deterministic core decides when to reach out and verifies whether each nudge worked by matching events from the other apps; a ReAct agent writes every message. Go ×3 services, Kafka, OpenRouter, PostgreSQL.
phone-focus Android screen-time collector written almost entirely in Go: usage snapshots every 10 minutes as CloudEvents, with an offline queue. Go (gomobile), Kotlin, React/TS.
Pet project · Jul 2026 – now · agent-os
An agent daemon on top of stock Ubuntu installs packages, manages services, edits configs and finds files by meaning (RAG on pgvector). Safety is structural, not confirmational: undo and a trash bin are prepared before every action, /etc and package state are snapshotted, the journal is append-only with a SHA-256 hash chain, the agent runs in an empty network namespace with Landlock. Local model by default (Ollama), cloud switchable by a word; the machine boots straight into conversation.
Stack: Go (five daemons under systemd), PostgreSQL + pgvector, Ollama / OpenRouter, React.
Team project (2 people) · Jun 2025 · auto-rag
Upload documents (PDF, DOCX, XLSX, images with OCR), get semantic search and context-grounded answers in a chat, tune RAG configs — top-k, relevance threshold, temperature, custom prompts — and share documents and configs between users. My part is the entire web application: Blazor Server on Clean Architecture with MVVM, EF Core over PostgreSQL/pgvector, MinIO object storage, auth, chat and sharing logic.
Stack: C#/.NET 8, Blazor Server, MudBlazor, EF Core, PostgreSQL + pgvector, MinIO; Python FastAPI microservices.
Team project · Oct – Dec 2024 · ai-news-assistent
News parsed from Telegram channels into MySQL, FAISS retrieval, GPT-4o answers served through a Telegram bot, a Gradio UI and a REST API. My part: the bot, the parser, the Gradio app and the whole RAG manager.
Stack: Python, MySQL, FAISS, GPT-4o, Gradio.
AutoML · research notebook · Dec 2025 Kaggle Playground S4E7: LightAutoML baselines against hand-tuned CatBoost and LightGBM with Optuna; best submission 0.895 ROC AUC.
DeepScriptum · team project · Dec 2024 – Jan 2025 Any document → Markdown via GPT-4o OCR run as a map-reduce pipeline; FastAPI + Celery + PostgreSQL. My part: the frontend.
aith-containers · course (ITMO) · Dec 2024 A three-service RAG application deployed to minikube from hand-written Kubernetes manifests.
hse-python-backend · course (HSE) · Oct 2024 A math API on raw ASGI, a FastAPI shop, WebSocket chat, Prometheus and Grafana monitoring.
ITMO_X5_HACKATHON · hackathon · Sep 2024 Retail tech-support chatbot: a 4-bit quantized LLaMA 3.1 8B over an E5 retriever.
AI_news_bot · pet project · Jul 2024 My first RAG project: a Telegram news-digest bot on GPT-4o and MS SQL Server.
courseWorkData · university coursework · Dec 2022 Clinic management software on ASP.NET Core MVC with EF Core and PostgreSQL.