Taking a one‑teacher yoga school off manual work.

Parveen Nair teaches embodied yoga from Rishikesh to a paying audience on Patreon, with a small volunteer team. Every weekly class meant pulling a 4GB file off the camera, compressing it, writing a title and description, numbering it into the right series, uploading it and telling the team. Teachers asking to join the directory arrived by email. Social posting was by hand. None of it was difficult; all of it was recurring, and it competed with teaching.

The rule is that autonomy follows reversibility. Filing a class is private and easy to undo, so nobody checks it. Publishing a stranger's photograph and biography on the client's site is public and hard to take back, so it waits for one click from one person. Posting to the paying audience waits for a person too. Payment is the strictest: the only thing that can grant a membership is an event signed by Stripe, because a browser arriving at a success page proves it came back, not that anyone paid.

Patreon keeps around a tenth of every payment before processing fees, owns the viewing experience, and has no way to represent a school with several teachers. Kula is the alternative: each school lives at its own address with its own price and colour, members subscribe to a school rather than to a platform, and a membership at one school opens nothing at another. Video streams through the server rather than from a shareable link, so access is checked on every request, including every time someone skips ahead.

It is checked the way it would be in production. A script walks the whole membership lifecycle against the live site through real systems: a real checkout session, a real test subscription, a webhook that rejects a forged signature and accepts a real one, a real sign-in email read back from the mail provider, the paid class streaming for the member and refused for a stranger, then a cancellation that Stripe itself signs and delivers, revoking access. All 13 checks pass. Stripe's hosted checkout was also driven by hand with the test card to confirm the payer comes back signed in.

Members can describe what they want to practise in their own words, such as humming bee breath, or tight hips from sitting all day, and get the three classes that fit best. Each class is cut into passages, its notes plus overlapping windows of what the teacher said in the recording, 634 in all, embedded with Workers AI into a Vectorize index. The best passage per class decides the ranking. A small model then writes one line per class on why it fits, using only that class's notes and the passage that matched.

The model explains a ranking it did not make, so a weak sentence costs a sentence, not a wrong class. Its output is tidied in code rather than trusted to the prompt. Transcripts never leave the server whole: members see a 28-word snippet of the passage that matched, visitors see none. And a daily ceiling counted in the database bounds what a public demo can spend on models, whatever the account's plan.

I direct coding agents rather than typing most of the code myself. Here that was Claude Code, with me owning the client conversations, the architecture, which failures were worth a guard, and the call on every trade-off. The working rule is that nothing is done until it has been checked against the running system, which is why the figures on this page are measured rather than estimated.