// systems I build & run
My Work
Systems I've designed, built, and run. Some are my own infrastructure and R&D — the place where the ideas behind the work get proven, on real hardware, in production, behind this very domain.
Selected systems
lazarev.cloud — Self-Hosted Production Platform
My own infrastructure, and my proving ground: a 9-node Proxmox cluster with NVIDIA GPU and Ryzen AI NPU passthrough for local ML workloads, provisioned end to end as code with OpenTofu. Defense-in-depth with a Vault-issued internal PKI and SSO across every service, a fully self-hosted CI/CD supply chain (GitLab, Harbor, Nexus, Renovate), and Prometheus/Grafana observability with automated 3-2-1 backups. Reproducible, reviewable, and operated end to end by me.
Stack: Proxmox · OpenTofu · Vault · GitLab · Harbor · Nexus · CrowdSec · Prometheus/Grafana · Pocket ID
Local LLM Agent Memory Stack
A production-grade external memory system for LLM agents — vector storage and retrieval with reranking, persistent agent memory, and a fast cache layer. Built and benchmarked against current research, running entirely on local hardware.
Stack: Qdrant · Mem0 · Valkey · Qwen3 embeddings + reranker
Local AI Media Pipeline
A ComfyUI-based image/video generation pipeline tuned for AMD Ryzen AI hardware (ROCm), with a fully autonomous build/setup flow. Large generative models without a cloud bill.
Stack: ComfyUI · ROCm · AMD Ryzen AI MAX+ 395
Edge & IoT
ESP32 environmental monitoring with live dashboards, and a Raspberry Pi security camera with servo control wired into n8n automation. Small systems, fully owned.
Stack: ESP32 · Raspberry Pi · n8n
Public work
AI without the fear — a course for beginners
I wrote this for my parents, then made it public. Five short lessons on using AI assistants safely, for people who find them confusing or intimidating: what a chatbot actually is and where it fails, which privacy settings to change, how to ask a question that gets a useful answer, why models state wrong facts with total confidence, and how voice cloning and deepfakes are being used against exactly this audience. Every figure is sourced to an official publication. English and Russian, plus a print-ready handbook and a slide deck for anyone who wants to run a session.