Online & async RL infrastructure

The algorithm is only part of the work. The rest is infrastructure: sandboxes that run hundreds of thousands of agent rollouts, and a pipeline that keeps the training accelerators saturated while it learns from live traffic without breaking it.

This role explicitly includes building the eval harnesses, sandboxed environments and async RL pipelines that research needs. The interview, and the onsite in particular, may reward reasoning about throughput, staleness and safety as concretely as about advantages.