Multiverse. Operations tooling for safer learning agents.

Multiverse treats reinforcement learning as an operations problem: not just training an agent, but deciding when a run is safe, measurable, repeatable, and ready to promote.

The repo includes custom environments, an agent registry, rollout infrastructure, memory indexing and retrieval, benchmark validation, and command-line tools for inspecting runs.

The core loop starts with a registered environment and agent, runs rollouts through a safety wrapper, records events, saves useful memory, and exposes the run through CLI tools. Instead of treating each experiment as a one-off script, Multiverse gives training runs a repeatable path for review.

The memory layer retrieves prior episodes and validation artifacts, which makes the project more than a toy RL trainer. A run can use earlier experience as context, while benchmark gates and safety checks decide whether a candidate is better.

The repo also documents cleanup work: large files were broken into support modules, runtime entrypoints moved under tools, and test coverage was organized around active behavior. That makes it a useful portfolio piece for engineering maturity, not just ML curiosity.

Multiverse is the project I would use to show that I can work with machine-learning systems as software systems in Python: registries, configs, CLI workflows, test suites, observability, safety boundaries, and performance measurements.

The best framing is not "I trained an RL model." It is "I built the tools around learning agents so experiments can be inspected, compared, and promoted carefully."