Fana

RESEARCH

How Fana works.

A simple explanation of what we build, and why it holds up.

01

Start with one job.

Take any job a company does. Closing the books at the end of the month is a good one. To a person it feels like one task. Underneath, it is a long chain of small steps: find the right invoice, match it to the payment, notice the one that doesn't fit.

We break every job down to steps that small.

02

Give each step a company.

A step only means something inside a real business. So we build the business around it: its books, its bank statements, its contracts, its habits.

Each step is built inside many different kinds of company, from small firms to nonprofits, and each one keeps its records its own way. That is how one job becomes more than a thousand worlds.

03

Build it the same way every time.

Every world is built by code from a single number, called a seed. Give it the same seed and you get the same world, down to the last cent.

That matters more than it sounds. Anyone can check a result, and a test you run today can be run again next year and still mean the same thing.

04

Let the world check the answer.

When a model finishes, nobody reads its work and gives an opinion. The world checks it. The books balance or they don't. The right invoice was found or it wasn't.

The world also looks at how the model got there, so a lucky guess doesn't count. Because every check is arithmetic, it gives the same verdict every time, and it can't be talked into a better score.

Every world is built by code, not written by another AI. Nothing in it is made up, and nothing about it depends on a model you don't own.

05

More than 1,000 worlds, and counting.

Every mark is a world a model can learn in: one step of one job, inside one company, with a problem waiting to be found.

06

Where we are going.

We started with finance because every number has to be right, which makes it the hardest place to fake an answer.

Coding is next. After that, the method doesn't care about the field: any job that can be broken into steps can be built this way, and that is where we are going.

A model of your own, trained on your own data.

Get in touch