The AI system I use to run my work solo

I work alone. No assistant, no operations team. Yet every day I still have dozens of small tasks to do: note-taking, answering messages, saving documents, monitoring servers, writing content, downloading and processing videos. Instead of hiring someone, I built an AI system to handle the repetitive work. This post describes how that system is running, so you can see how one person working alone can organize AI.

Main idea: divide roles like a small team

I don't use one chatbot for everything. I divide the system into "team members", each doing one clear job:

• Coordination brain running on Mac: where data lives, automated processes and the gateway to AI models.
• Executive arm running on a Windows machine: an AI agent that takes tasks from the brain, manipulates the browser, downloads and extracts videos, replies via Telegram.
• Knowledge vault is an Obsidian vault: strategy, processes, daily logs are all markdown files that both I and agents can read.

The two machines connect via a private network, so the agent on one can call services on the other without opening ports to the internet.

Automated workflows: n8n is the backbone

Most repetitive work goes through n8n, a drag-and-drop automation tool I installed myself. Each workflow is a chain of steps: receive data, process it, call AI if needed, then write the result back in the right place.

Agents also send tasks to each other through a webhook on n8n. Each message has the same format: who is sending, sending to whom, task type (assign, return result, alert, status update) and content. Because of this common format, I can add new agents without rewriting how they communicate.

Three-layer memory

The biggest weakness of chatbots is forgetting. I solve it with three layers of memory:

1. Short-term memory for events, habits, frequently used settings. Agents quickly write and read back.
2. Knowledge database (PostgreSQL with semantic search) for long documents, scripts, large knowledge. Documents are cut into chunks and converted to vectors so AI finds the right chunk.
3. Obsidian vault for things I need to read and edit by hand: strategy, work processes, daily logs.

Every time an agent starts a work session, it reads a common status file first, then works. When done, it writes what it did. So the next agent knows where the previous one stopped.

One gateway for many AI models

I'm not locked to one provider. All AI calls go through a routing gateway (I use 9Router), so I swap models for each task without editing each workflow. Work needing deep reasoning uses a strong model, simple work uses a cheap model.

This website is also part of the system

The page you're reading, the course page and store are all built and run by me with AI agents. Posts, tool library and prompts are managed in the admin panel. Agents can read and suggest content via MCP, but everything agents write goes into a review queue before going live.

What I learned

• Start from one task, not the whole system. This system grew gradually from very small workflows.
• Write everything down. Status files and daily logs are what keep multiple agents working in sync.
• Human reviews the last step. AI is fast, but posting, sending or deleting still goes through my hands.

If you're also working alone and want to start automating, my next post covers how to pick the first task to give to AI.