I built a JARVIS, and it told me off

"Nothing has moved on any of your projects in five hours."

Correct. I was asleep. My dashboard has a British accent, and strong feelings about my sleeping habits.

That's roughly the state of the thing I spent the better part of last week building: useful, faintly judgmental, and not finished. A few people asked what it actually is after I posted a video of it, so here it is, for better or worse.

It started as a boring dashboard

The first spec was dull on purpose. I've got around thirty projects running under Claude Code at any given time. I wanted one screen telling me which were alive, which were stuck waiting on me, and what I was spending on tokens, because that number can get away from you faster than you'd think.

Then I decided I wanted to be able to talk to it, and have it talk to me. Not type at it. I type all day. If I'm going to ask "what needs me today," I want to ask the way I'd ask a person, out loud, while doing something else.

That pushed it out of dashboard territory. A screen you read is a report. A thing you ask is an assistant, and an assistant has to know things a report never has to, and a good assistant anticipates your needs rather than just spits facts back at you.

So the spec kept growing. Email went in, because half of what needs me arrives there. Then the agents I already had went in with it, because they were the reason there was so much to keep track of in the first place. Perry runs travel: flights, visa deadlines, the ninety-day itinerary. (Perry has been right about a date I was wrong about, which I'm still a little sour about.) Moya isn't one agent but a whole pod of them, with Pilot running the queue. Seldon is the product pipeline, from spotting an opportunity through planning, building and launching it.

Each of those was useful alone. What none of them had was a front door. I was the front door. I was the thing holding in my head that Perry had a question, Moya had finished something, and a project had been sitting blocked since Tuesday. Not very efficient.

JARVIS was a butler first

The first build was pushy, reporting my idle hours back at me like a timesheet. A personality problem. Easy fix. Now it says "morning has rather come and gone" and gets on with it. It even makes the occasional sarcastic remark, which I appreciate. Behavior change by editing a text file. Thirty seconds, no code, no deploy. First time in my career that editing a document has fixed somebody's tone or level of competency.

The accent is intentional. JARVIS stands for "Just A Rather Very Intelligent System", and the name came from Edwin Jarvis, the butler who worked for Howard Stark long before anyone wrote a line of code for him. Everybody who has ever seen Iron Man knows what I mean. I wanted my own, and it's the thing people actually picture when they say JARVIS: spoken, faintly formal, coming from somewhere in the room. Only my 'Intelligent' System is named George.

Most of the work is underneath

Some people will look at it and think it's cool. Some will think I wasted a week on a toy (to be clear, I did other things in parallel). Some will find the design is ugly and the UI doesn't work for them. All fair, and none of it was or will ever be a factor in what I built for myself.

This is my Fleet, built to my aesthetic, and that layer is wide open. Anyone doing this can go as far as they like in whatever direction they like, because the look is cosmetic. It matters enormously to the person using it and it's a small share of the actual work. The work is underneath: a stable foundation, a toolset, code that behaves tomorrow the way it did today.

Most of the time went into PRDs and architecture documentation. During the build there was nearly as much auditing as building, which sounds tedious because it is — for the agent, at least. Less so for me. Formal, structured audit steps are part of all my builds. Aeryn takes care of that, and in exchange she gets regular upgrades.

Which is why the questions I judge my tool on are all first person. Does it work for me. Does it keep me organized. Does it make me want to open it in the morning.

And then the layer I'd just called easy broke. I flipped a switch to turn on a feature. The switch worked. But visually, the toggle didn't change, so the interface told me it hadn't turned on. About ten minutes to find that the display was throwing the update away, and 13 lines to fix it. The cosmetic layer had spent a morning insisting a feature was off while it was quietly on. Not critical, but not nothing either.

Yes, it's another AI post

I know. People hear "AI" and their eyes glaze over, and honestly, most days mine do too. A good share of the hype is produced by people who have never had to make one of these things work, with real data, for a business that has to keep running while they do it. They browse X, listen to 'creators' who tell them they automated their entire marketing team with AI, then turn around and demand that to happen in their company. It can be infuriating, but is it substantially different from only a couple of years ago, when it wasn't AI but the newest, shiniest SaaS tool that was to 'save' the company?

Sitting out AI isn't a strategy either, and neither is arguing with the fact that it is coming and fast. The hardest part about adopting AI and AI based systems is people and process. It needs guidance and supervision. It isn't creative. It isn't free of bias. And it screws up, which is a strange thing to hold against it, since really so does every human being I've ever worked with, and yes, I am including myself.

No amount of hype or hate changes any of that, in either direction.

Nobody needs one of these

If your work fits in one place and you can see all of it at once, this solves a problem you don't have.

However, if you have multiple workflows and touchpoints across tech or teams, having one will grow on you faster than you'd expect. It did for me. Within days I stopped digging through my inboxes and checking usage windows on the AI companies' settings pages. It's all right there, and if not, I can ask and will (more likely than not) be answered.

You don't have to build it yourself either. Someone can build one on your stack, for the way you work, match a design you like, and it'll fit better than anything off a shelf, because the fit is the entire point. Remember, people and process. There's a shorter version of all this on the work page, including the honest bit about what it actually takes.

Would I build it again? Yes, and I'm not finished. Databases, uptime and the n8n orchestration go in next: the parts of the stack that currently only tell me they're unhappy by failing in front of a customer.

It's coming to Southeast Asia

In a few weeks this thing gets on a plane with me. Working, seeing some of the world. I'll share what George and I will be working on and where.

I'm one person with a lot of moving parts and I'm about to have no fixed address for a while. Which is either the reason I built it or a very elaborate way of avoiding the question of why one person needs thirty projects. The latter is a problem, believe me I know.

Ask me in December how it all went, or if you can't wait, follow me, I will be sharing regular updates.