About Refine

A spreadsheet error in one of the most influential economics papers of the decade helped justify austerity across Europe and the US. Millions lived with the consequences. The error was only found years later, by sheer luck, long after the damage was done.

Flawed analysis shapes policy, investment and lending every day, costs billions every year, and mostly goes unnoticed. AI makes it worse: rigorous-looking work is now cheap to produce.

Refine exists to catch these errors before anyone acts on them. We're building the infrastructure that verifies high-stakes work; think CI/CD for everything beyond code. It reviews a document like a team of experts, checking the maths, data, logic and argument, in minutes.

The best AI review system. Refine 5 beats frontier models such as GPT-6 Astra in over 90% of blind head-to-heads across ten fields, from economics to biology. It combines the top frontier models at hundreds of times normal compute, so no one model's blind spots get through. We tune it on documents seeded with the errors frontier models miss.

The de facto AI reviewer for economics. Econometrica and the AER run every paper they publish through Refine. It's used at 24 central banks and by researchers at Oxford, MIT, Stanford, Yale and Cambridge; Cambridge, UCLA and NYU provide it to theirs.

Beyond academia. 29,000+ reviews sold at 1,000+ institutions, including litigation consultancies, hedge funds and companies bidding on public tenders.

Enterprise-grade. ISO 27001 certified, SOC 2 Type 1 (Type 2 in progress). On the web and inside ChatGPT, Claude and Codex.

Where we are. Seed stage: fewer than 10 people, $5–10M raised, and revenue in the $1–10M range.

The role

You'll build Refine with a swarm of AI agents: they write most of the code, and you do the thinking. You own work end to end, from the customer who asked for it, to the code running in production, to the numbers that say whether it worked.

The work covers the whole product:

The core AI pipeline. The review engine itself: new checks, how frontier models are combined, and what happens when one of them fails.

Organizations and B2B. Self-serve organizations, team and department plans, institutional checkout, and what large clients need next, such as single sign-on.

Pricing. New plans, packs, discounts and trials, end to end through payments.

Integrations. Refine already runs inside ChatGPT, Claude and Codex, and there's more to connect.

You'll make the everyday technical calls yourself, so the founders can spend their time on customers and growth. You'll work with Nacho García, who leads engineering, and with the founders, Yann Calvo Lopez (CEO) and Ben Golub (Chief Scientist, formerly a tenured economics professor at Harvard). Our advisors include OpenAI co-founder John Schulman.

How we build

This is the part we're proudest of. The codebase is built for people and agents to work on together, and it holds up.

Cloud native, everything in code. We run on managed cloud services, and every environment is infrastructure as code in Pulumi: production, an EU stack, and a throwaway stack for any branch.

Domain-driven design and hexagonal architecture. Business rules live in a domain layer with no infrastructure in it. Aggregates raise their own domain events, and long workflows run as durable sagas. Architecture tests enforce the boundaries, so nobody, human or agent, quietly breaks them.

A real testing pyramid. Unit tests at the base. An acceptance test for every use case, against a real database. Whole review pipelines run end to end with scripted models. Browser end-to-end and accessibility tests at the top. Every pull request gets a security scan and an AI review before a person looks at it.

Measured before it ships. We benchmark changes to models, prompts and checks against the current pipeline and against one call to the strongest frontier model, then roll them out behind feature flags.

Built for LLMs that fail. A review makes many calls to frontier models from several providers. Workflows survive crashes and pick up where they stopped. Every call and stage has its own time budget. Stages degrade gracefully, with fallbacks. If a review can't be completed, the customer gets their credits back automatically.