Why we build software that never sees your data

When we started building Apeirogon, the easiest path would have been the one everyone else takes: ship a thin client, run the real work on our servers, and bill by usage. It's a proven model. It's also one that requires us to see everything you do.

We didn't want that trade-off. Every file you open, every formula you write, every draft you throw away — none of it should have to pass through infrastructure we control just to get useful AI assistance. So we built the other way: the model runs on your machine, the data stays on your machine, and the product works whether or not you're online.

This wasn't a marketing decision first. It was an architecture decision that happened to produce a marketing story. Once you commit to on-device inference, a lot of other choices follow naturally — no usage metering, no rate limits tied to server capacity, no outage when our infrastructure has a bad day.

The harder part was making local inference feel as fast and capable as a cloud round-trip. That meant real investment in model efficiency, not just wrapping an API call. We think the result is worth it: software that works for you, not on you.