What changed by Friday.

The problem

Zealandia counts its little spotted kiwi the hard way: catch, band, measure, release, and record it all in three spreadsheets. Turning those into answers — how many birds, are they surviving, are they in condition to breed, will the population still be viable in fifty years — was a specialist job across half a dozen tools. Each study took about a week and cost the sanctuary around NZ$5,000 to commission, so it was only run when it absolutely had to be.

What we built

KiwiPop: a password-protected web app where staff upload the same three workbooks and get the whole study back in minutes — population estimates, body condition, survival, breeding, movement mapping, a fifty-year viability model, and an avian-flu stress test across eight species. Built with Claude Code by one person, in the hours between other work. Along the way it audited the sanctuary's raw Excel files and found that a locale round-trip had silently swapped day and month on every ambiguous date, turned an 80-hour survey into 1,472 hours, and mangled sixteen microchip numbers. It wrote the parser that reverses the damage, and handed back repaired workbooks with every fix highlighted.

What changed

The NZ$5,000 fee is gone. Reruns cost nothing, so the analysis happens whenever new captures come in rather than once a year. Questions like "what if we translocate fifteen birds every three years?" get answered in the meeting where they're asked.

The question that mattered most was one nobody had been able to answer. With avian flu spreading around the world but not yet in New Zealand, the sanctuary could model its impact before it arrived — and see, in hard numbers, how much extra mortality the kiwi population could absorb and still hold. For a team that has spent decades rebuilding this population, that was more than a chart: it was evidence that their most iconic species has a breeding population that can survive a shock, and that the conservation effort as a whole is working.

And because a conservation organisation's species records are sensitive, the product itself contains no generative AI at all — the maths is classical statistics, deterministic and reproducible, and the data never leaves the sanctuary's own tool.

The problem

Every reading of every bill is on the record in Hansard, but it's scattered across two government websites, written in parliamentary procedure, and the corrected version lags by about nine months. There's no votes API. Most votes are party votes, so any tool that says "your MP voted X" is usually making it up. And a civic tool read by partisan audiences has to be right every time.

What we built

A platform where you vote on real bills in plain English and see how Parliament actually voted — with a regional map of what New Zealanders said, next to what the House did. Behind it, an engine that reads Hansard and links every vote to its bill with provenance attached, and a daily pipeline that scans for new bills, harvests the votes, drafts a neutral summary of each, and posts to social media — all on schedule, with no one at the keyboard. Every summary the public reads has been reviewed and approved by a person before it goes live.

What changed

One person built and now runs a complete civic product covering the current Parliament — 329 plain-English bill summaries, verified voting records for every reading, and a social channel that publishes daily on its own. New Zealanders cast more than 9,200 votes on it in its first 45 days, with no advertising spend. The daily operation needs one human step: reviewing the summaries that appear in the morning. Everything else runs itself and reports back. Provisional results are labelled as provisional and corrected automatically when the official record catches up, which is how the platform has published outcomes a week ahead of Parliament's own API.