Hello, world. We're round3.
We started a company to answer the hardest question in production AI: is this change actually better?
For the past few months, I've been quietly working on something new. Today, I can finally say a little more about it: we've started a company. It's called round3.
The deeper we got into production AI, the more one problem kept showing up. Teams can change a prompt, swap a model, or modify an agent in minutes. Knowing whether any of those changes actually made the product better is remarkably difficult.
A new model can look stronger on a public benchmark. A new prompt can feel better in a demo. An agent can handle the happy path beautifully. Then it meets real users: messy inputs, multi-turn conversations, latency expectations, cost constraints, and edge cases with no generic benchmark captured. At that point, too many consequential decisions still come down to a vibe check.
Offline evaluations are part of the answer, but not the whole answer. They help teams test changes against scenarios that resemble their own product and identify which candidates deserve a closer look. Production is where you learn whether an apparent improvement changes the outcome your users and your business actually care about.
That's the problem we started round3 to solve.
Our first product is three.dev. It shows you how your AI agent fails and helps you find the best configuration for your use case. From there, it turns your production data and your team's judgment into a specialized model for that exact job, built to beat the frontier model it replaces on quality, latency, and cost.
We believe every team should be able to create and own the AI its products depend on.
This isn't a product launch. We're early, working closely with a small number of teams, learning from real AI systems operating under real constraints, and building around the problems we find. The work is already producing stories worth telling, and we'll share the first of them soon.
I'm incredibly fortunate to be building this with Ximo Guanter, my co-founder and CTO; Andrea Moscatelli, Dani Rodríguez Hernández, and Raúl Moreno Salinas, our engineers; and Bryan Lee, our product lead.
They are exceptional at what they do, generous with one another, and a lot of fun to build with. Starting a company inevitably comes with plenty of uncertainty, and I feel very lucky to be going through it with this group.
If you're building an AI product and have found yourself asking, "Is this actually better?", I'd love to compare notes.
More soon.