Atelier
I once raised a $1M seed round to start a software company. I hired a team of sixteen, built a product, took it to market, and heard from our first paying customer about twelve months later. I thought that was fast. Today I've shipped ten products in parallel for a few hundred dollars, and it keeps getting cheaper and easier.
Managing a team of bots, and changing how they work as the tools change, is a new skill for me. People now call this a software factory. I was inspired by the workshops I saw in Paris, so I call mine an atelier. This page is a public record of my experiments and what I'm learning from them.
In the future, AI will be private, free, and everywhere.
Today most AI costs money every time you use it, and it means handing your data to someone else's server. I expect both to fade. Features people pay for now will run on their own machines at no cost. Chrome has started shipping AI models inside the browser, so I built two extensions to see how far that goes.
Mirlo
A Chrome extension that swaps words in a news article into the language you're learning.
I wanted to grow my Spanish vocabulary, so I built it on Chrome's built-in Translator API. I read the news with it every day.
Skimly
A Chrome extension that summarizes just the section you're skimming, and expands it back with a click.
Summaries of a whole article never captured its essence for me. When I skim, I want to compress the dense sections and keep reading the rest.
In the future, agents will find, try, and buy software.
Agents can now build much of the software people used to buy, on demand. What they can't easily make is maintained data. Go-to-market is different when the buyer is an agent. So I built two feeds for agents rather than people. An agent reads the free page, and a metered trial on the MCP server is the path to paying.
SunBiz Signals
Every new Florida business from the last seven days, shaped for an agent to read alongside a CRM.
An insurance agent pastes it into Claude with their client list and skips an hour of prospecting.
Destination Pulse
Leadership changes at US tourism boards, so agencies can reach a new leader before the next vendor review.
The same recipe in a new industry. Machines detect the changes, AI checks them, and I approve.
In the future, we'll manage agents like seasoned employees.
My early AI products treated the model like an intern: I told it exactly what to do and checked everything it handed back. As models improved, I started letting them decide, and coaching them in Slack when they got it wrong. A seasoned employee keeps notes on what they've learned. These agents rewrite their own instructions and commit them to git, so I can read every change and roll one back.
The Abba Times
A one-page newspaper for my daughter, written by an agent every school day.
The editor rides its desks for news, weather, sports, and jokes until the day's paper fits on one sheet, in time for her lunchbox before she leaves at 6:40.
Curated Austin Events
A weekly calendar of Austin events on technology, philosophy, Buddhism, and science, picked by an agent.
Each week it goes looking for events I'd want to attend. When it misses, I tell it in Slack, and the next week's picks change.
In the future, every niche will have its own magazine.
A niche magazine collects what's happening in its corner, writes about what it means, and keeps it current. That used to take a staff, so most small niches never had one. Now a newsroom of agents can do the work. Researchers comb the sources for anything new, writers turn what they find into listings and stories, and an editor shapes each week's issue.
Zen Retreats
A magazine of meditation retreats across North America, with new retreats each week and an editorial by theme.
I used to check the local zen centers for retreats, and the listings were sparse when they existed at all. It was one of my first projects, and the one I learned the most from.
Cloudflare Apps
A trade publication for open-source apps you can deploy to your own Cloudflare account, with a directory and practical guides.
The same masthead on a different stack. I built it to learn Cloudflare's own infrastructure, from queues and workflows to the open-source models it hosts.
In the future, apps will be personal.
Software has been built for the many, because only the many could pay for it. When building gets cheap, an app can serve one person or one family and answer to them, not to an engagement chart. These two started at home.
Alfee
Simple reading games for people with aphasia, with audio support and evidence-informed exercises.
I built it when my mom was in speech therapy after a stroke.
Good Friction
Small tools that put a pause back into social feeds, starting with LinkedIn.
Paginated feeds and slow page loads used to give us a moment to choose. Endless feeds took it away. Good Friction puts it back.