What I build with, and why each one is here.

Drafting, extraction, agentic work

The most reliable at long, multi-step work that has to not drift. Where a job has twenty dependent steps, this is what I reach for.

Classification, bulk processing

Cheap and fast enough to run over thousands of items. Used it to triage newsletters and resolve three thousand job addresses.

Anything that must stay local

gpt-oss and Qwen run on hardware you own. A generation behind frontier, and completely adequate for drafting and document Q&A.

Speech

Voice interfaces where typing is the wrong input: field work, hands busy, driving between sites.

Every application

One framework for the interface and the server. Pages render as static HTML where they can, which is why these sites load instantly.

Interface

Not a preference so much as the default that everything else assumes. Fighting it costs more than it saves.

All of it

Catches the class of mistake that otherwise reaches a client's screen. Non-negotiable on anything someone else relies on.

Database & auth

Postgres with the tedious parts done. Row-level security means access rules live in the database rather than in application code you can forget.

Data work

Anything involving a few thousand rows, a geocoder, or a training run. The right tool for work that isn't a web page.

Automation

Self-hostable, so the same workflow runs in someone's office or in the cloud without a rewrite. That portability is the whole point.

Hosting

Deploys from a git push and serves static pages from the edge. No servers to patch, which matters when there's one of me.

The on-premise stack

How an entire system gets installed identically on every client machine, and rolled back when an update misbehaves.

Accounting integrations

Where most NZ small-business financial data actually lives. Bookkeeping automation starts and ends here.

Writing the code

Worth naming because it's the honest answer. AI-assisted editing is how I ship at this pace alongside a full-time degree.