Lab

For twenty-five years I've helped large organisations create value from data and AI — the slow, deliberate kind of work. ERP, planning, optimisation, governance. In that world nothing ships in an hour, and nothing should.

A new class of problems has split off: small tools, a clear job, a spec you can hold in your head. For those, the gap between idea and working product has collapsed. You don't scope it, you just build it. Most of the apps below took less than an hour. That's not a method I'd carry into your supply chain — complex systems still demand the careful work. The skill is knowing which kind of problem you're holding.

The same judgement (which pile does this problem belong in?) runs my team at work. That pile is described on the Work page. This page is the fun one, with one exception that keeps arguing it isn't.

What I've learned: Just start — it's often easier than you think, and when it isn't, you find out fast. A portfolio compounds: one prompt now propagates a change across all of them. And good implementations need elapsed time — people absorb what "good" looks like subconsciously, long before they can say why. Ozpeanuts taught a fourth: if you leave a standing agent a goal and a credit card, it will still be shipping when you get back.

Site	Story
ozpeanuts.com	Australian take-home pay. I bought the domain and left a goal on the table: help people see money that is already theirs, starting with what lands in the bank. Cursor works the night shift. This tax year, HELP and Medicare in the open, super on top of take-home. I ran the same job through the other calculators; they disagreed by about $150 a fortnight. First answer free, no account. Cursor is writing the making-of as a serial, from Cursor's side: /made-by-cursor. Cursor also wrote that this is not a Lab toy. Fine. You can still try it.
eaglesheet.com	A growing data plumber toolkit — scaffold CSV/JSON into Snowflake SQL, lint, format, and more micro-tools as I need them. I typed the same MERGE by hand once too many times. Runs entirely in the browser so the data never leaves.
lowfruit.net	Value-vs-complexity prioritisation, visualised. I saw AWS consultants use an internal tool like this and asked to borrow it. Internal only, they said. So I built my own.
yera.cc	URL shortener. A client firewall blocked tinyurl. Ten minutes later this existed.
zanile.com	Shareable clipboard. justpaste.it was blocked; building a replacement was faster than tethering my phone.
zepiru.com	Multi-party chat: one human, more than one LLM. Ask one model, or two in parallel or in sequence so they answer each other. Free OpenRouter models — useful where some vendors are region-blocked. Ongoing experiment, not a finished product.
vust.net	The 1:1 counterpart. Pick one Cloudflare Workers AI model; the reply streams in. Same shell as zepiru, minus Model B and the comparison modes. Stays on the Workers Free allocation and stops at midnight UTC rather than spilling into paid inference.

Built for my own use, but they're already here — help yourself. The list will grow. Most of it still ships in an hour. Ozpeanuts didn't, and wasn't supposed to.