SDE 3 · CAW · OPEN TO ROLES
05 — HOMELAB

My side-server
that thinks for itself.

Where I prototype agentic AI in the open. A self-hosted stack of containers — orchestrated with Portainer, exposed through a VPN + reverse proxy, and wired up with n8n & Claude.

WHY A HOMELAB

I learn by building real systems at home.

My homelab is where the AI thesis becomes practice. Multiple Docker containers managed through Portainer, wrapped behind a VPN and reverse proxy, and stitched into agentic workflows with n8n + Claude. It's small, but it's running 24/7 — and every weekend, it gets a new capability.

I wanted a sandbox where I could break things, ship things, and watch my own systems run — the homelab is exactly that.— Hemant
Containers
12+
Uptime
99.4%
Orchestrator
Portainer
Agents
n8n + Claude
ARCHITECTURE

From the public internet to a thinking agent.

Every request flows the same way — through the VPN/proxy, into Portainer-managed containers, and (where it matters) up into the agentic AI layer that decides what to do next.

Internet

Public
HTTPS

VPN + Proxy

WireGuard
Nginx PM

Portainer

Containers
under control

Containers

App, DB,
queue, cache

n8n

Workflow
orchestrator

Claude + MCP

Reasoning +
custom tools

RUNNING NOW

A snapshot of what's online.

LIVE · {{ updated daily }}
Portainer CERUNNING

The single pane of glass. Manages every container in the lab — health, logs, stacks, volumes — without ever touching the CLI.

// orchestration · web ui
Nginx Proxy ManagerRUNNING

Reverse proxy in front of every service. TLS auto-renewed, internal subdomains, access lists. The lab is internet-facing, but only on my terms.

// tls · ingress · routing
WireGuard VPNRUNNING

Zero-trust ingress. Admin surfaces (Portainer, n8n, Grafana) are only reachable via VPN — never exposed to the open web.

// vpn · zero-trust
n8nAGENTIC

My agent orchestrator. Multi-step workflows chaining Claude calls, MCP tools, webhooks, schedules, and human-in-the-loop steps. 14 flows running.

// agentic · workflows · webhooks
MCP ServersAGENTIC

Custom Model Context Protocol servers expose real tools (DB, files, search, internal APIs) to Claude. Each one is small, typed, and tested.

// claude tools · stdio + http
Claude ProxyAGENTIC

A thin gateway in front of the Anthropic API — adds auth, logging, rate limits, and a unified retry/streaming layer for everything in the lab.

// llm gateway
Postgres + RedisRUNNING

Stateful backbone. Postgres holds app data & n8n state; Redis handles caching and the agent queue. Boring, on purpose.

// data · queue · cache
Grafana + PrometheusRUNNING

If it runs in the lab, it's scraped. Dashboards for container health, agent latency, and the bills my LLM proxy is racking up.

// observability
Watchtower + VaultwardenRUNNING

Watchtower auto-pulls image updates on a schedule. Vaultwarden keeps all the lab secrets self-hosted and out of any cloud vault.

// updates · secrets