Finance systems · Applied AI
I build automation finance teams can trust, in tax and procure-to-pay systems and in my own AI projects.
Live productAI-led build
An infinite-zoom drawing app: zoom into any detail and keep drawing inside it.
It’s also my test of how far AI can take complex software. I direct and review the work, and agents write all of the code: about 165,000 lines so far.
Interactive demoAgent research system
What would it take to make a smartphone from nothing?
Agents research the parts, tools and industries behind it, and code checks every change before it joins a shared map of 1,507 records.
Interactive demoUX prototypes
Six working UX prototypes, built with AI agents.
Each one tries a different way to explore the smartphone research, so I could choose a design before committing to a full learning site.
Interactive demoAgent skill
An agent workflow that turns a room description or reference picture into a structured floor plan and an editable Sweet Home 3D project.
Every inferred dimension is marked, so you can see what was measured and what was guessed.
Lets a Claude agent choose when to compact its own context, so multi-week tasks continue without supervision.
A home-ideas notebook I built with AI and use on my phone to save photos, links and notes.
A search tool for job listings: an AI researcher gathers openings, and each result stays linked to its original posting so it can be checked.
Small experiments I used to inspect real browser behavior and compare tradeoffs before choosing an approach for the app.
Compare two ways of turning a pen stroke into a clean, editable shape.
Find how many circular arcs a smooth stroke actually needs.
Compare when to calculate a stroke’s outline: while drawing, after the pen lifts or a little of both.
Investigate thin lines that disappear while zooming in an infinite canvas.
Compare ways to draw accurate curves when zoom produces enormous coordinates.
Compare selection indicators for objects too small or too large for an ordinary box.
Give AI a useful role. Make its output reviewable.
Set clear formats, keep one source of truth and make relationships explicit, so AI output can be checked and reused.
Keep what AI proposed separate from what a person approved, and a prototype separate from a proven result.
Measure behavior in the real environment, check every change before it’s accepted, and be clear about what a test does and doesn’t prove.
Give agents saved progress and control over when they pause, so long tasks need less supervision.