An applied AI research company.
Perception — also written Perception Club, after perception.club, where the work is published — builds reinforcement-learning environments for AI agents. An environment is the tasks an agent practises on, the tools it acts through, and a verifier that scores what it actually produced. The agent improves from its own outcomes rather than from imitating examples, and every gain is measured on work it has never seen.
What we work on
An agent that investigates, decides and acts does not get better by reading examples of the work. Our central programme is building the environment it practises in, and certifying what came out: a gain on held-out tasks, and no quiet regression in whatever the use case depends on. How environments work covers the method, the anatomy of an environment, and the process we follow with a use case.
Recent work includes agents that learn house rules from reward alone with the model frozen, agentic investigation over a data warehouse, teams of agents against single agents, and a studio that generates an environment from a customer's own warehouse. We publish results as research notes, including the runs that did not work.
Earlier research
Before environments, we studied whether a trained model can be repaired without retraining: small supervised patches, and a certificate that decides whether a patch is safe to install. The findings — that patches leak into neighbouring capabilities, that they destroy each other when merged, and that imitation-style evaluation overstates what a model can actually do — are the case for training on outcomes instead. That work stays published at earlier research.
The name
Perception and Perception Club are the same company; the second is how the domain reads. It is not the activity club in Chennai, the talk show, or the open-source juggling club of the same name, and the AI-twin product once advertised under the name is retired. The company's work today is the research published on this site.
Get in touch
Tell us the job you want an agent to do well, and what doing it well means. We'll tell you plainly whether an RL environment is the right tool, and what building one would involve. Write to contact@perception.club, or bring us a use case.