Work
A look at what I build, what I am experimenting with, and the technologies I keep coming back to.
An AI-first personal operating system
Atlas is the project I keep returning to. It pulls memory, daily work and the way I actually live into one system, built on the same standards I hold enterprise software to. The question underneath it is simple enough to say and hard to answer: can software help you think, or does it only help you look busy?
Status: being built, slowly and on purpose
Selected work
Notes, tasks and reference material live in a different app each, and none of them know what the others hold. The goal was one system that remembers what you know and can use it to help you decide what to do next.
One user, no team, no runway - it is built in the hours around a full-time job. It holds personal data, so the trust bar is higher than any demo needs. And it has to stay useful on the days I have twenty minutes, not two hours.
Built narrow and deep rather than broad: fewer features, each one load-bearing. Retrieval is favoured over generation wherever the honest answer is a thing I already wrote. Shipping speed was traded away for a data model I will not have to unpick later.
In active development, deliberately unreleased. The part that works today is memory: it answers from what I have actually written rather than inventing a plausible reply, which was the whole point.
Areas of focus
Putting LLMs, prompt design and intelligent workflows inside enterprise software. The hard part is never the model. It is everything around it: the edge cases, the trust, the moment a user asks it something you never planned for.
Nine years of frontend architecture and business platforms that run in production for people who did not choose to use them. That constraint shapes everything about how I build.
Chat and automation work, where interaction design runs straight into engineering limits. Getting a conversation to feel natural is mostly a design problem with an engineering bill attached.
How I think
None of these are original. They are the things nine years of shipping kept proving, usually by charging me for the lesson first.
A demo has one path through it and a friendly audience. Production has neither. The work that decides whether software is any good happens after the impressive part is finished.
Every clever mechanism is a thing someone has to hold in their head at 4pm on a bad day. The systems that survive are the ones a new engineer can be wrong about safely.
Nobody outside the team has ever been impressed by a diagram. They notice whether the thing was fast, whether it was right, and whether it did what they expected. That is the entire scoreboard.
Most software adds choices and calls it flexibility. The better move is usually to make the sensible thing the default and let the exception cost something.
The person using this is busy, half distracted and mid-way through something else. Software written for an attentive user is software written for a user who does not exist.
Everything else
Years of articles on AI and product engineering, plus the personal writing that has nothing to do with either.
Read the archive →Weekly conversations about growth, healing and the mindset behind a life you actually want.
Listen →Experiments, half-finished ideas and the things that never made it as far as a landing page.
View GitHub →