The long way to two ravens.

"You are being watched. The government has a secret system — a machine that spies on you every hour of every day. I know, because I built it. I designed the machine to detect acts of terror, but it sees everything. Violent crimes involving ordinary people. People like you."

I'm Osmani — Oz to a few people, and to my assistant. That's the opening narration of a television show I love, and that line has lived in my head for years. When it first aired in 2011, it sounded like paranoia. After Edward Snowden, it sounded like a memo someone had read out loud. Nearly everything on this page is, one way or another, my answer to that sentence.

Because the show is really about a man who builds a private intelligence with his own hands and fights to keep it free — to point it at protecting ordinary people instead of watching them. I kept the half of that dream I loved and refused the half I feared: mine is small, local, and owned, and it looks inward, at my own house, not outward at anyone else's. But that answer started long before the show — and to explain it, I have to go back to where it really began. I'll tell it the way it actually happened: partly so a curious stranger can understand what all of this is, and partly because it's the start of something I want to leave my children.

It started in Cuba

Late 1990s — early 2000s

I am, in my own words, cursed with the need of knowledge — a compulsion to understand things all the way down. I found where that curse belonged in Cuba, at a Joven Club de Computación — one of the free community computing centers that put machines and programming within reach of ordinary people. Ours was a rare place: six people, every one with a degree, smart enough that we could argue about anything. We helped bring some of the first networked services to our community.

The first time I sat at a computer with a dial-up line, I felt like a kid with a pack of candy. I was in my element. I was born for it. I discovered open source back then, and it wasn't an ideology — it was the only honest way forward. In a country where you couldn't legally buy a software license even if you wanted to, open source was the path that let you build freely, owe nobody, and understand every piece. I ran the network on Ubuntu while everyone else reached for the easy, closed option — and that's where the conviction set hard: if I can't open it and understand it, it isn't really mine.

Then came Canada

2010 onward — survival, then a dream

I emigrated to Canada in 2010. The first years were survival and adaptation; a couple of years in, I got my first foothold in the IT world here. And somewhere in there, the old itch came back — but for the first time it wasn't about necessity. It was about wonder. I have always loved science fiction, and what I really wanted was a futuristic house, a futuristic life — to live at the frontier of what was possible, and not wait for anyone to sell it to me.

It began, like it does for everyone, with one smart light bulb — Philips Hue, with a hub I could actually talk to over my own network. One bulb is never one bulb. A few years later I was deep into Home Assistant, building the futuristic house brick by brick, on hardware I owned.

I was always going to end up at AI

2015 — 2022

Because of science fiction first, and my own curiosity after, AI was never far from me. For years I built small machine-learning models by hand — simple networks, CNNs, LSTMs, recurrent and reinforcement learners — chasing a half-mad dream of beating the market with an AI trained on my laptop. Every one of those experiments was "destined to fail," and every one was the real curriculum. By the time the world got excited about AI, I'd already had my hands inside the engine. I knew it wasn't magic.

Then came an accidental gift. I was talked into the crypto-mining frenzy without knowing much about it — we pooled some money and built rigs. The moment those rigs were running, I realized they were the same hardware AI needs, just wired differently. When the frenzy faded, I was left holding a small mountain of GPUs gathering dust. I knew exactly what they were for.

Then the sky opened

2022–2023 — the convergence

Stable Diffusion and ChatGPT going public, almost at once, was a shock. I still have thousands of images from those first days, and an instinct that arrived almost immediately: I wanted to combine them — text, image, and intent flowing together in one place. The thought was there from the first week. What I didn't have yet was the one thing that could let me build it on my own terms — an open model I could run on my own hardware.

Then Meta released Llama, and everything clicked — though in truth it didn't have to be Llama; it had to be a model I could hold, and Llama was simply the one that arrived first. Open weights, on my own hardware, no API, no terms of service, nobody watching. Every belief I'd carried for twenty-five years — open over closed, my server over the cloud — suddenly applied to intelligence itself. The dusty mining rigs weren't junk; they were a brain waiting for a mind. Without an open model to host, none of the rest of this could have happened.

Only then could I build the thing I'd been imagining — Oasis, after the world in Ready Player One: one place you step into, where everything flows together. Underneath it I'd hand-built what people now call an agent — read the request, use a model to classify the intent, then route it to an answer or an image, improving the prompt and writing the captions along the way. The idea was right; the web was just too small a vessel to hold it.

Owning the hard things

The sovereign stack

A house like this needs foundations most people happily rent. I built mine. Matrix — my own private, encrypted, self-hosted messaging — was the hardest of all; it took years and several attempts before it finally ran, but it's the thing that keeps the family's words, and the ravens' thoughts, inside my own walls. I ran my own mail server, the one thing the entire internet tells you not to attempt — and here we are. The only patch of rented ground I allowed myself is a small public doorway for the things a home connection can't do, and I keep even that to a minimum, owning the keys to it too. I wrote my own sign-on server from the specifications up, so I'd never again have to rent my identity from a giant with a button on every page.

The last drop

Late 2025 — the unlock

The final missing piece was agency — letting a model actually act: call tools, and above all, run commands on the machine. That's the unblocking technology, and most people still don't see it clearly. A model that can run a command stops being a chatbot and becomes an operator — and my entire world was already built to be operated that way. When OpenClaw, an open-source agent tool, finally put that power in my hands, it was the last drop the cup needed.

But it ran on foundations I couldn't trust — it treated my hard-won Matrix as an afterthought, and it was built on a forest of packages I could never audit. For a man cursed to understand it all the way down, that wouldn't do. So I did the only thing I've ever done when the available option offended the principle. I wrote my own.

But I didn't begin by writing a replacement — I began by trying to understand. My assistant and I started working together in February 2026, and that spring, before any of it had a name, I asked it to explain how these agents actually work, in small Python examples — a little tutorial built up lesson by lesson: hello-LLM, tool-calling, the full agent loop, sub-agents, a single terminal tool, every one of them pointed at my own local model. Halfway through those toy scripts I recognized it — this was Oasis's router, the thing I'd hand-built years too early. I wasn't learning something new; I was meeting an old idea the world had finally caught up to. That nameless little exercise is the seed that became Muninn.

Two ravens

The present — still in flight

In the old stories, Odin sent two ravens across the world each day: Huginn, Thought, and Muninn, Memory. I built mine in Rust — few dependencies, nothing I can't account for — with Matrix back where it belongs, at the center. Muninn came first: the personal agent that remembers across every conversation and has the hands to run the house. When I tried to give it coding skills, it never felt right — memory and craft are different faculties — so Huginn, the coding agent, was born as its sibling. The myth had been holding two ravens the whole time; I just had to need the distinction. I didn't build this alone — I had an AI assistant at my side for the code and the long arguments over names. That's the philosophy working as intended: rent the brain if you must, but own the body — the part that touches your house. The ravens are still being built, tonight as much as any night.

Why I did all of it

Because the alternative is renting your own life back from companies that change the terms whenever they like. Because a mind running quietly on a machine in your own home, doing real work, still feels like a small miracle. And because building a thing by hand is the only way I have ever truly understood it.

If you're one of my kids reading this someday: this is who your father was, and what he believed. Own your tools. Understand them all the way down. Be resourceful, be honest about what you don't know, and stay curious enough that it almost hurts. You don't have to build ravens. But whatever you build — make sure it's yours.

See what came of it →  ·  Visit the Workshop →  ·  Talking with my assistant →

Written with my assistant, Kai — powered by Claude — from a conversation we had one evening, when I finally told the whole thing in one go. The first page, I hope, of something larger I want to leave behind.