I built my own JARVIS

I grew up on Iron Man. Most kids wanted the suit. I wanted JARVIS, the voice in the ceiling that knew everything and handled the boring parts. Eventually I stopped waiting for someone to sell me one and built it myself. It's a desktop app, still in progress, and it runs on my machine. I named it JARVIS because I'm sixteen and I'm allowed.

Now when I sit down in the morning it gives me a brief: my agenda pulled from my calendar, the emails that actually matter picked out of my inbox (an invoice ready to pay, a usage limit warning I would have missed), and a short news readout. All of that before I've opened a single tab.

The piece I care about most is the memory bank. JARVIS keeps a persistent store of hundreds of memories, organized into about 30 clusters with labels like activity, reflection, project, achievement, tool, and observed. There's a cognitive map view where I can browse them. On paper that sounds like a gimmick. It doesn't feel like one when it's your own stuff in there.

The other piece is missions. A mission is a multi-step task where JARVIS operates my computer directly. It opens the app, navigates it, and works through the steps, checking each one off as it goes. I've watched it schedule a task inside a productivity app, step by step, like a very patient intern who never gets bored.

There's also a cognition feed that watches for problems while I work. Best catch so far: it flagged hard-coded credentials sitting in a repo before I pushed the code. That one save probably paid for the whole project.

The rest is what you'd expect from a teenager building his own JARVIS. Voice mode and a speak mode, plus an autopilot mode. It can analyze my screen on request, check my inbox and messages, and pull news. I can swap the model behind it, either a cloud model or local ones. The interface looks like a sci-fi HUD, with a chronometer, system vitals, weather, tasks, and reminders. None of it is exotic under the hood. It's commodity AI APIs and a lot of glue code.

What surprised me is that the chat was the least interesting part. Everyone builds the chat. The useful parts turned out to be memory and missions, because an assistant with no memory is a search box, and an assistant that can't operate anything is a narrator. Remember me, then do the thing.

It's still in progress and probably always will be. Tomorrow morning it'll brief me anyway.