Plunax LabsEnvironments where agents learn.
Plunax Labs is a small research team building reinforcement-learning environments and the data analysis that tells you what an agent actually learned.
Two problems we spend our time on
Reinforcement-learning environments
We design and build environments for training and evaluating agents: clear observation and action spaces, reward functions that measure what you meant, and tasks that stay hard as agents get better.
- Custom simulators and Gymnasium-compatible environments
- Reward design and reward-hacking audits
- Evaluation suites and reproducible baselines
Data analytics research
Training runs, logs and production data hide most of the interesting answers. We turn them into analysis people can make decisions with.
- Training-run and agent-behaviour analysis
- Experiment design and statistical evaluation
- Data pipelines and dashboards for research teams
How we work
Measure first
Every environment ships with the metrics and baselines that show whether it is learnable, and how.
Small and reproducible
Seeds, configs and versions are recorded, so any result can be run again by someone else.
Read the behaviour
A rising reward curve is a question, not an answer. We look at what the agent actually does.
Working on agents or data?
Tell us about the problem. We reply to every message.