Recurvia is a deep reinforcement learning lab building AI-powered infrastructure and instrumentation which accelerate the development of advanced artificial intelligence systems.
Built by researchers and engineers from
Educated at
Our mission is to build the intelligent infrastructure and instrumentation that enables teams of any size to accelerate the development of advanced AI systems. We believe Reinforcement Learning will be the central learning paradigm powering both the next generation of intelligent systems and the infrastructure that enables them. That’s the future this team was built for.
Model sizes and compute requirements are increasing exponentially. More GPUs is one answer. More efficient is a better one.
Developing advanced AI systems increasingly depends on iteration velocity and compute efficiency, at every level of the stack. These are problems with search spaces far too large and too high-dimensional to explore by hand, and the decisions arrive faster than any human can make them effectively. The next advances in AI efficiency will therefore be powered by AI itself. We are building a new generation of AI-powered infrastructure for everyone training serious models, from start-ups to frontier labs.
Demand for compute continues to outrun both supply and progress in hardware. The industry's answer is often simply more: more accelerators, more data centres, more power. We prefer more efficient.
Every wasted cycle has a price, and the hyperparameters that drive a run are still chosen up front by brute-force multi-run search, then left fixed while conditions change. Our solution is a policy trained to intelligently adapt hyperparameters continuously inside a single run - observably and within the limits its human operators set.
Paramorph is a foundation model trained from the ground up with multi-agent RL to optimise the training processes of other models. Developing it demanded observability at a scale and granularity existing experiment trackers were never intended for - so we built Metrana. Now, we're making it available to select lab partners.
Metrana began life as the tooling we needed to develop Paramorph: metric logging at frontier scale; reinforcement learning workloads as first class citizens; AI-native analysis and autonomous experimentation. Now, we're making it available to select lab partners.
Paramorph is the first of a new class of models: neural networks that learn to make other neural networks more efficient and performant. It adapts learning strategies live during training, continuously adjusting hyperparameters and training behaviour to improve convergence, scalability, and compute efficiency. Itself a foundation model trained with multi-agent reinforcement learning, Paramorph treats optimisation as an intelligent, continuously evolving process rather than a static configuration problem.
.webp)
As the training horizons and model sizes grow, small inefficiencies translate into millions of dollars in wasted compute. The labs that learn fastest will be the ones running on AI infrastructure powered by AI, and trained with reinforcement learning.
BACKED BY LEADING DEEP-TECH INVESTORS AND ACCELERATORS
The next generation of AI will be built and optimised with RL-powered infrastructure.
We are building it.
