I work in AI strategy at a large bank. I spend my days figuring out how 7,000 people should adopt AI tools. And then I come home and build my own.

This isn’t a contradiction. It’s the same instinct: I don’t trust anything I haven’t built (or at least taken apart).

The setup

The core is n8n — an open-source workflow automation tool running on my home server. On top of it, I’ve built what I call the Family Bot: an orchestrator that connects WhatsApp and Telegram to a set of sub-agents, each handling a different domain.

The architecture looks like this:

  • Main Orchestrator — receives messages, classifies intent, routes to the right sub-agent
  • To-do Agent — reads and writes to my Notion task database
  • Internet Search Agent — answers questions using web search
  • Document Reader — retrieves files, guides, and reference material from a Notion knowledge base
  • Image Generator — creates images on demand and sends them back to the chat

Each sub-agent is a separate n8n workflow, called by the orchestrator as a tool. They share a common pattern: webhook trigger → process → respond. Error handling flows to a dedicated catcher workflow that logs failures and notifies me.

Why not just use ChatGPT?

Because ChatGPT doesn’t know where my kid’s passport is. It doesn’t know my Notion task list. It can’t send a reminder to the family WhatsApp group. It can’t pull up the insurance policy PDF when I need the claim number.

The value of a personal AI assistant isn’t intelligence — it’s context. It knows my stuff. And because I built it, I know exactly what it can and can’t do.

What’s next

The current system handles text well. The next layer is:

  1. Voice — speech-to-text input, text-to-speech output. Talk to the bot instead of typing.
  2. Memory — RAG over personal documents using Qdrant vector database. Three collections: family docs, personal notes, and work references.
  3. Calendar awareness — proactive suggestions based on what’s coming up, not just reactive responses.

The goal isn’t to build a product. It’s to build something that makes my family’s daily life slightly less chaotic. And to understand — by building — what these systems can actually do.

Because that’s the thing about AI strategy: you can read all the Gartner reports you want, but you don’t know until you’ve debugged a webhook at 11pm because the Telegram API changed its rate limits.