Gubbins

Every AI assistant you use starts from nothing, so you explain your business again, and again, to each of them. Gubbins holds that context once. It learns a small business, or a life full of admin, through a conversation, keeps what it learns in a memory you can see, correct, archive and export, and lends slices of it to the assistants you choose, under grants you set, with a receipt for every read.

It does no work of its own. Doing the job is the assistant's business, and a bank that also did the work would be competing with the very things it needs to serve. There's no company behind it, which is deliberate. Every decision is mine, and I get to watch what breaks.

Gubbins started as the opposite thing. It did your admin, one earned step at a time, with a board of jobs you approved. In August 2026 I took all of that out.

Three things pointed the same way. Testers kept asking for control over what the thing knew about them rather than for more jobs it could do. The interoperability item near the bottom of the roadmap was more interesting than the product it was attached to. And a neutral bank you own, which serves every assistant and competes with none of them, is the one thing no assistant vendor can structurally offer, because they all want to be the place your context lives.

I could have kept both halves. I decided one product manager cannot carry two products, and that a hedge would have got me a worse version of each.

The old half did not come out when I made the decision. It came out when the new half had visibly worked, which is a different day and the whole point. I wrote the proof as the condition: a grant issued, a recall served and receipted, a fact deposited, reviewed and accepted, then the grant revoked and the next read refused, all watched happening end to end.

That ran on 24 August. The jobs code came out on the 25th. The tables underneath it were frozen rather than dropped, because someone's export should keep returning their own rows even after the feature that made them is gone.

The obvious promise for a product like this is that it redacts what is sensitive before anything leaves. I refused to make it. One missed name breaks a promise like that permanently, and no scrubber catches every name.

What it promises instead is scope and evidence. You granted these areas and no others, and here is the receipt for every read. That is a promise the system can keep on its worst day. Scrubbing still happens, and it is called best effort everywhere it appears, which is all it can honestly be called.

One product manager who doesn't write code, with AI doing the building, on a schedule and mostly unattended. Supabase for the database and sign-in, Vercel for hosting, GitHub for the code, and Claude Code tying the three together and doing the work. There's one planning column that only I fill, and an empty column means nothing happens that day.

The part I didn't expect to matter is the writing. Long technical implementation plans, a vision document and a business plan, all written properly before anything gets built, because a model builds what the plan says and nothing it doesn't. Everything the project knows lives in the code repository, including what's broken.

Under 18s are turned away at sign-up and again at the front door, and that refusal is a fixed sentence from ordinary code rather than something the model decides. A rule a model can be talked out of isn't a rule.

Nothing reaches an assistant that the owner has not granted, and the bank refuses rather than guesses when a request falls outside a grant. Sensitivity is marked in the same green, amber and red vocabulary the storage rules already used, rather than inventing a second ladder for people to learn.