I architect and ship end-to-end AI systems - from RAG pipelines and multi-agent orchestration to production backends and infrastructure.

My focus is on building things that work reliably at scale: retrieval systems, LLM integrations, intelligent assistants, backend machinery, and production deployment.

• LLM Engineering - designing and building multi-agent systems, orchestrating tool calls, and optimizing context for production workloads.
• RAG & Retrieval - architecting end-to-end search pipelines with hybrid retrieval, reranking, and vector DB optimization.
• Backend Engineering - building async Python backends with clean architecture, auth, and integrations.
• Bots & Automation - shipping production-grade AI assistants for Telegram and WhatsApp with multilingual support and knowledge bases.
• DevOps - containerizing, deploying, and monitoring services with CI/CD and observability.

• Building enterprise-level AI assistants with RAG, memory, and multi-step reasoning.
• Architecture of modular toolstores, agent routers, prompt stores, and model abstractions.
• Designing LLM-first backends with clean boundaries and predictable tool execution.
• Adaptive pipelines for production: ingestion, validation, observability, evaluation.

• Primary: Python 3.12 (async, typing, Pydantic-first)
• Data & Queries: SQL, JSON, YAML, TOML
• Web & Scripting: JavaScript / TypeScript, HTML, CSS
• Systems & Automation: Bash, PowerShell, Dockerfile, Makefile
• Exploring: Go, Rust, C, C++ (reading-level)

• OpenAI API, Anthropic API, Google Gemini, Grok, DeepSeek
• Open-source LLMs
• Ollama - local model deployment
• Embedding models: text-embedding-3-large, SentenceTransformers, multilingual
• LangChain, LangGraph
• Qdrant, ChromaDB, FAISS
• HuggingFace, tiktoken
• Cross-encoder reranking (BAAI/bge-reranker-v2-m3, Jina AI, .etc)
• Langfuse - LLM observability
• Prompt engineering (structured output, function calling, tool calling)
• Multi-provider architecture (OpenAI, Anthropic, Gemini, DeepSeek, Groq)
• SSE Streaming - real-time response streaming
• Document parsing: PDF (pypdf), DOCX (python-docx), TXT, JSON, .etc

• FastAPI, Uvicorn (ASGI)
• Pydantic v2 - data validation, settings management
• Starlette - middleware (rate limiting, security)
• AsyncIO, SQLAlchemy 2.0
• Alembic (migrations)
• PostgreSQL, psycopg 3 + psycopg-pool
• MongoDB
• Redis
• httpx, aiohttp - async HTTP clients
• Multiprocessing & multithreading
• JWT auth (python-jose) - access + refresh tokens
• OAuth 2.0 (Google)
• Webhook systems (Telegram, WhatsApp Cloud API)
• OpenAPI / Swagger UI / ReDoc

• aiogram - Telegram Bot API framework
• aiogram FSM - finite state machines for multi-step dialogs

• RAG (Retrieval-Augmented Generation), Hierarchical RAG
• Hybrid Search - Dense + Sparse (BM25)
• PostgreSQL Full-Text Search (tsvector, GIN indexes)
• Reciprocal Rank Fusion (RRF)
• Cross-Encoder Reranking

• Apache Bench, JMeter - load testing
• pytest, pytest-cov, pytest-asyncio, pytest-mock
• FastAPI TestClient / httpx

• Docker / Compose (dev, prod configs)
• GitHub Actions
• Ubuntu Server, systemd
• Health Checks (/health, /ready, /metrics)

• Structured logging (JSON/text, rotation)
• Custom metrics (latency, cache hit/miss, error rates)
• Psutil - CPU/RAM monitoring
• SMTP/webhook alerts

• Prompt Injection Protection - custom multilingual guardrails (regex filters, input sanitization, context boundary enforcement)
• LLM output validation, content filtering, safe tool execution
• Rate Limiting (per-user/IP via Redis)
• Retry with Exponential Backoff

• Ruff - linter & formatter

• Flowise - no-code LLM orchestration
• Chroma Cloud
• n8n

Feel free to reach out:

• Email: skvozdymperegara@proton.me

"AI systems are built, not summoned. Good engineering beats magic every time."