Python Developer · Backend Engineer · AI Engineer
I build backend systems where AI agents are architectural components — not API wrappers. Currently a Python developer at a startup working on industrial automation and AI.
AI-powered backend systems and agent architecture. The model is one component inside a larger system — bounded by deterministic logic, connected to tools and knowledge sources, wired into real business processes.
Deterministic planning, LLM generation. Control flow and formally checkable rules live in code; the model handles only what genuinely requires reasoning.
Agents as platform components. Stateful graphs with human-in-the-loop interrupts, parallel fan-out and custom reducers for concurrent writes — shipped with auth, integration contracts and offline tests, not as standalone scripts.
Tools and MCP instead of direct coupling. A production MCP server with JWT auth and scoped tokens serves a nested knowledge agent; the calling agent never touches the protocol.
Hard real-time budgets. Concurrent graphs under a fixed per-turn limit with fallback replies, semaphore-limited model concurrency, field-level Redis merges between parallel graphs.
| Backend & Data Design | Async services, layered structure, multi-tenant models, explicit state machines, migration discipline |
| System Architecture | Multi-service decomposition, service boundaries, service-to-service auth, graceful degradation |
| Agent Architecture | Graph design, explicit state, interrupts and resumption, tool interfaces, agent authorization |
| AI Engineering | Prompting under anti-hallucination constraints, structured outputs, provider fallback, latency and cost budgets |
| Retrieval & Knowledge | Vector search with metadata filtering and LLM reranking, exposed to agents through tools |
| Async Processing | Task queues with specialized worker pools, distributed locks, pub/sub, progress streaming |
| Integrations & Automation | CRM, telephony, speech and storage providers; webhooks; end-to-end automated pipelines |
| Testing & Delivery | Black-box API suites, offline graph tests, container stacks, monitoring, CI/CD |
| Technical Ownership | End-to-end from schema to deployment; mentoring developers; integration guides and deploy runbooks |
Core · Python 3.12 · FastAPI · SQLAlchemy 2.0 (async) · Pydantic v2 · Alembic · pytest
Agents & LLM · LangGraph · LangChain · MCP / FastMCP · tool calling · structured outputs
Retrieval · Qdrant · ChromaDB · embeddings · metadata filtering · LLM reranking
Models & inference · Whisper / faster-whisper · GigaAM · YOLOv8 · PyTorch · vLLM · OpenAI API
Data & Queues · PostgreSQL · Redis · S3 / MinIO · TaskIQ · Celery · ARQ · RabbitMQ
Infrastructure · Docker & Compose · nginx · Grafana · Loki · Prometheus · GitHub Actions
Integrations · Bitrix24 · amoCRM · LiveKit (WebRTC / SIP) · ElevenLabs · Telegram Bot API
Secondary · Next.js · TypeScript — internal consoles and dashboards
The recurring shape across my projects: a thin API layer, background workers doing the heavy work, and an agent runtime whose state is explicit, observable and testable offline.
Commercial projects, mostly in private repositories — described without client-identifying details.
Real-time voice AI agent. Live inbound phone calls. Three concurrent graphs: the dialogue agent, a checker running while the caller is still speaking, and a background worker for profile extraction and knowledge lookups. Deterministic script planning, 9-second turn budget, self-hosted Russian ASR. 11 services, 667 offline tests.
Document compliance agent. Validates official documents against a formal regulation. One graph turns the regulation into per-section rules offline; a second runs structural checks in code and semantic checks as parallel LLM calls, streaming progress live.
Call analysis platform with MCP. Scores call recordings pulled from CRM with a two-tier agent — a scoring agent that delegates knowledge questions to a nested knowledge agent, which reaches the knowledge base through a production MCP server with a forwarded token chain.
Conversational requirements agent. 18-node graph with human-in-the-loop interrupts, parallel fan-out, and an LLM resolver mapping free text onto available choices. Paired with a FastAPI platform: role-based access, Celery workers, service-to-service auth.
Industrial defect detection. Four-model YOLOv8 cascade on GPU, from an industrial camera feed through to PLC actuation of a physical reject mechanism.
- Agent orchestration — coordination, recovery and state across multi-graph systems
- MCP as the integration layer between agents and business systems
- Retrieval architecture — ingestion pipelines, hybrid vector + keyword search
- Making AI systems observable, testable and reproducible in production
Open to conversations about agent architecture, LLM-driven backends and AI automation.