Company / What we believe
These are not aspirations. They are the constraints we use when deciding what to build and what to leave out.
We build the model, the training pipeline and the product.
It is slower, and it is the only way to stand behind what we ship. When Atom gets something wrong we can trace it back to data, training or serving and fix it, rather than filing a ticket with a vendor.
It also keeps our incentives straight. We are not reselling someone else's capacity, so we have no reason to push you towards the most expensive path through the product.
The most useful assistant is one you can be candid with.
Features that trade privacy for convenience do not make the cut. Conversations are not used for training, deletion is real, and sharing is opt-in and revocable.
We publish our sub-processors and hosting locations so you can check our claims instead of taking them on trust.
Reasoning, citations, code and output are visible.
Trust should be earned in the open, not assumed. When Atom thinks, you can read the thinking. When it searches, the sources are one click away. When it runs code, you see what ran.
The same applies to us. We publish our evaluation settings and sample sizes, and report regressions as plainly as gains.
A focused team, shipping steadily.
We would rather do a few things properly than everything approximately. That shows up as a product with fewer switches, clearer defaults and release notes you can actually read.
It also means we say no often, and try to say why.
If we fall short of any of these, tell us.