I'm building my entire AI infrastructure in public. Here's why.

Six months ago I quit my job at Microsoft and started building.

Not a startup. Not a SaaS product. I started building the personal infrastructure I wished I had: AI assistants that actually know my context, automation workflows that connect everything I use, and a knowledge system that doesn't forget what I did last Tuesday.

Along the way I've burned through $55/month on a single AI assistant that could have cost $5. I've rebuilt my entire note-taking system three times. I've wired together Claude, Obsidian, n8n, Telegram, pgvector, and a half-dozen other tools into something that mostly works and occasionally does something that genuinely surprises me.

I haven't written about any of it.

That changes now.

What this blog is

This is a build log. Real systems, real failures, real decisions.

I'm an AI-first builder. I use Claude Code with multiple agents running simultaneously, Obsidian as an operating system, OpenBrain as a shared memory layer across all my AI tools, and a Telegram bot named Maya that lets me pick up work from the beach. None of this is theoretical. All of it is running, breaking, and getting rebuilt.

Every post here will be rooted in something I actually built, fixed, or learned. No "5 ways AI will change your workflow" listicles. No hypothetical architecture diagrams. If I'm writing about it, I've shipped it or I've failed at it. Both are worth documenting.

What's coming

I have three series planned.

Personal AI Infrastructure covers the engineering behind my multi-agent personal system. The first post is about how my AI assistant was costing me $55/month and how I redesigned it. After that: shared memory with OpenBrain, cost circuit breakers, and presence-aware AI that knows when I'm away.

Alfred OS is the story of replacing Notion, Todoist, and a dozen other apps with a structured knowledge system in Obsidian. Twenty record types, everything linked, every AI session auto-writing back to the knowledge base. This is where the ADHD brain meets systems engineering.

AI-First Workflows covers the practical side. How I use Claude Code for task management. How four different AI tools share context without forgetting each other. When to run models locally and when to call the cloud.

Why build in public

Two reasons.

First, I learn better when I write things down. Explaining a system forces you to understand it. Every time I've written up a project, I've found at least one thing I should have done differently.

Second, I want to work with people who build like this. I spent years at Microsoft shipping products to 100M+ education users. Now I help organizations navigate AI adoption through consulting and workshops. The best conversations I have start with "I read your post about X and we're dealing with something similar."

That's the goal. Show the work. Start the conversation.

How this site works

This blog runs on the same Next.js site as my portfolio at sandtorv.ai. Posts are plain markdown files with validated frontmatter, statically generated, no CMS. The blog section itself was built in a single Claude Code session, which felt like the right way to launch a blog about building with AI.

Every post will also generate a LinkedIn post and an Instagram carousel. If the content is technical enough, maybe a thread. The distribution pipeline is part of the system I'm building, and I'll write about that too.

Get in touch

If you're building your own AI-augmented workflows, or if you're a hiring manager wondering what an AI-first builder actually does all day, I'd love to hear from you.

Find me on LinkedIn, GitHub, or email me at magnus@sandtorv.ai.

Let's build.