Zoa
Research

One Engine. Every Domain.

Most forecasting tools are built for one thing. Bloomberg's terminal knows bonds. OpenWeather knows rain. Your options desk quant knows implied vol. But the world doesn't divide itself neatly into silos - and the best predictions are made by people who can read signals from across disciplines.

That's the premise behind Zoa Research. The New York-based AI lab, founded by Greg Volynsky and Sam Damashek and backed by Y Combinator's Summer 2024 cohort, is building quantitative forecasting models that don't specialize. They generalize. One cross-domain engine that can be pointed at earnings surprises, earthquake risk, energy demand, or supply chain disruption - and produce calibrated, probabilistic predictions.

The bet isn't that AI can beat a domain expert at their own game. The bet is that a model trained across thousands of different event types will spot patterns that no single expert ever could.

What makes this more than a clever pitch is the mechanism. Zoa embeds large language models inside multi-agent optimization loops with fixed evaluation policies. The system builds models, tests them, and iterates - automatically. Internally, they describe it as "AlphaEvolve for forecasting." The AlphaGo parallel is intentional: you don't train a superhuman Go player by having them study games. You train them by having them play millions of games and learn from every loss.

Zoa is doing the same thing, except the game is predicting what happens next in the real world.

Founders rarely get their origin story right on the first try. Greg and Sam met at Carnegie Mellon in 2017 - not in a machine learning lab, but somewhere in the intersection of curiosity and ambition that CMU tends to produce. Greg went to Harvard Law. Sam went to Jane Street.

Together, they pulled in a team that includes a CMU CS PhD in automated reasoning (proof complexity, SAT solvers), a Princeton CS full-stack/ML engineer, and an operations lead. The team reads like a YC dream list, except they actually met each other before the application.

The same underlying architecture is used across domains that have nothing obvious in common. That's by design. A model that can switch between earthquake forecasting and energy demand isn't a generalist - it's an arbitrageur of pattern recognition.

Most research grants come from foundations with agendas, or corporations with NDAs. Zoa's approach is different. The Z-Grant program hands $10,000 to ML researchers - no strings attached beyond sharing what they learn and joining the research community.

It's a classic YC move: build the network before you need it. A lab that funds researchers creates a distributed brain trust of people who understand forecasting from first principles. When you need to hire or collaborate, you're not cold-calling - you're texting people you already funded.