Research Engineer - Post-Training

Pluralis Research works on Protocol Learning: training and serving large models in a fully decentralized way on small consumer-grade devices connected via the internet. Despite being dismissed as infeasible, we have made significant advances on this problem, most recently Agora, a permissionless run that pretrained an 8B model from scratch on consumer GPUs spread over the internet, with no single participant ever holding the full weights (tech report).

While many of the core research problems have been solved, Protocol Learning unlocks a series of new challenges. For the mission in full, read A Third Path:
Protocol Learning. Agora gave us a pretrained 8B model. Post-training is how we make it useful for agentic use-cases. But every post-training stack you've seen assumes a datacenter — synchronous rollouts, fast interconnects, trusted workers. Ours gets none of that. It has to run on consumer GPUs, and Macs spread across the public internet, training a model whose weights no single participant ever holds, with rollouts arriving from a geo-distributed inference pipeline at high latencies.

Your primary role is to make RL post-training work here anyway — the algorithms and the system, end-to-end.

• Build the post-training stack:
  You build the RL training loop end-to-end: rollout ingestion from the geo-distributed inference pipeline, reward computation, policy updates, and getting updated weights back out to the network. You set the direction, and you make things happen.
• Invent the algorithms:
  Standard RL recipes assume on-policy rollouts from fast, trusted hardware. You adapt them to asynchronous, high-latency, partially trusted generation: staleness tolerance, off-policy corrections, and communication-efficient policy updates.
• Ship first post-trained models:
  You build the evals that show the models are improving, and you take the first decentralized post-trained release from run to public artifact.

• Hands-on RL post-training:
  You've run RL post-training on large language models — RLHF, RLVR, or reasoning-focused RL — and touched the systems layer yourself: rollout generation, async training loops, weight synchronization. Not just launched jobs on someone else's stack.
• Strong engineering:
  Production-quality Python and PyTorch: concurrency, failure handling, profiling before optimizing.
• Research ability:
  Publications in RL post-training, asynchronous or distributed RL, or nearby fields are a strong signal. So is unpublished work you can defend in detail.
• Mission alignment:
  You believe Protocol Learning is the viable third path for collective, trustless, and sovereign AI.

• Experience training over slow networks, or with decentralized or federated setups.
• Familiarity with serving-engine internals such as vLLM or SGLang — our rollout pipeline is a serving system.
• Experience with reward modeling or building verifiable-reward datasets.
• Experience with P2P networking and NAT traversal.
• Experience at proprietary, open-weight and open-source AI labs

• Equity-Heavy Package:
  We offer significant ownership for key technical contributors in addition to a high base salary.
• Remote-First Culture:
  Flexible work environment with team members distributed globally.
• Visa Sponsorship:
  Optional full visa sponsorship and relocation support to either Australia or the US.
• Open Problems:
  Training and serving frontier models on hardware you don't control, over networks you don't own, mostly has no published answers yet. You'll write some of the first ones.

• We work remotely across the world, with the main teams in Australia and North America. You'll need to be comfortable working across timezones.
• Applicants must have professional-level English proficiency (written and spoken).
• Recruiters: we aren't looking for agency support at this time. We'll reach out if we need help.