Seventy percent is zero

There is a number that has been quietly deciding the future of analytics, and the number is seventy.

Ask a large language model a question about your data and it will answer correctly about seventy percent of the time. When people hear this, they tend to split into two camps. One camp says: incredible, the machine does most of an analyst's job. The other camp has worked as an analyst.

Because here is the thing nobody puts on the conference slide: an analyst who is right seventy percent of the time is not seventy percent of an analyst. They are zero percent of an analyst. Nobody ever paid for answers. Answers are cheap; the warehouse is full of them. People pay for the right to act on an answer without checking it. That right is the entire product. Strip out the trust and what remains is a very confident coin with a thirty percent chance of landing on your foot.

I had seven years to learn this the slow way. Five of them I spent in data engineering, building pipelines and warehouses, the plumbing nobody admires until the day it leaks. Then two years as a data architect, and those two years deserve their own sentence. I worked, at the worst stretches, twenty hours and more in a day. I do not recommend it, my calendar from that period should be studied as a cautionary document, and yet I cannot pretend it taught me nothing. When you sit inside that many companies' data problems for that long, the problems stop looking like incidents and start looking like a pattern.

The pattern goes like this. We would build a client a proper analytics stack. Clean models, fast dashboards, everything labeled. The client would thank us sincerely, admire the dashboards the way one admires a relative's new kitchen, and then, three weeks later, email me a question the dashboard already answered. At 23:40 someone somewhere always needed one more number, and the someone answering was me. The dashboards were not the product. They were furniture. What people actually wanted was to ask a question in their own words and get an answer they could stake a meeting on.

So the temptation arrives on schedule: let AI do it. And the temptation is real, I felt it too. But it comes with two locked doors. The first is that the companies most in need of an AI analyst are precisely the ones that cannot, or legally may not, hand their warehouse to somebody else's cloud chatbot. The second door is the number seventy.

Tvaras is what we built to open both doors. It is a self-hosted AI data analyst: it lives next to your data, not the other way around, and the whole architecture is organised around one obsession, turning seventy into a number you can sign your name under. Trustworthiness is not a feature on the roadmap. It is the roadmap.

I say "we", and the "we" matters more than the architecture.

Some companies begin in garages. As far as I can reconstruct, Tvaras began somewhere around a third steep of shu puerh. Matas is an old friend, and our friendship has been conducted substantially in gongfu tea: small cups, long sessions, the kind of unhurried format in which a bad idea has nowhere to hide for two hours straight. An idea that survives the whole pot is usually worth writing down. Tvaras survived a great many pots. Matas runs growth now, which means he spends his days persuading the world of something he first had to persuade me of over tea, and he is better at both than I would like to admit.

Aušrinė and I have known each other for more than ten years. How we met, and why the beginning was awkward, is a story I am not telling here. It is a good story. I am saving it, partly for dramatic effect and partly because I am still deciding which of us comes out of it looking worse. What matters is that ten years later the awkwardness has fully depreciated, we are grown-ups and, by every legal definition, real businessmen. Aušrinė runs operations, the discipline of making sure the rest of us are only allowed to break things that can be fixed. Every founding team needs exactly one person like this. We were lucky; we got her.

That is the honest origin. Not a flash of inspiration, but seven years of accumulated irritation, two locked doors, one stubborn number, a teapot, an old story I still owe you, and three people who decided that an answer you cannot trust is not an answer at all.

The dashboards, I am told, are still beautiful. The teapot gets more use.