Member of Technical Staff, RL & Compute Environments

What you’ll work on

• Design representative compute-environment abstractions: state, feasible placement actions, transition assumptions, rewards, and explicit constraint signals for heterogeneous workloads.
• Build reproducible offline evaluation and counterfactual-analysis workflows that compare candidate policies against baselines using held-out workload, architecture, and drift conditions.
• Develop policy-learning and planning experiments that use action-conditioned performance predictions while reporting calibration limits and uncertainty-sensitive failure cases.
• Define shadow-evaluation and narrowly scoped online-experiment protocols with predeclared metrics, stop conditions, and rollback criteria; partner with Runtime on execution but do not own the production placement path.
• Measure how policy quality changes with workload mix, hardware capability differences, delayed effects, and distribution shift, separating reward improvement from operational reliability.
• Work with Measurement & Data Infrastructure to specify the environment/action/outcome evidence needed for reproducible training and evaluation slices, without taking ownership of the underlying data platform.
• Communicate decision evidence and known limitations to World Models, Runtime, and Safety & Verification; independent controls—not this role’s reward or policy—determine whether an action is permitted.

What you bring

• Demonstrated work in reinforcement learning, contextual bandits, planning, or off-policy evaluation where state distributions change and decisions have delayed system effects.
• Ability to turn a systems problem into a falsifiable experimental environment, including action semantics, reward trade-offs, baselines, held-out conditions, and variance-aware measurement.
• Strong programming and experimental-systems practice: build reliable research pipelines, inspect failures, and make experiments repeatable rather than optimizing only a headline metric.
• Practical scheduling or resource-allocation intuition, including the trade-offs among latency, throughput, utilization, contention, and feasibility across heterogeneous resources.
• Sound judgment about safe empirical work: distinguish simulation or offline evidence from online evidence, set bounded experiment conditions, and state uncertainty plainly.
• Clear cross-functional communication with ML researchers, runtime engineers, data/measurement partners, and independent safety reviewers.

What progress looks like

• A documented compute-environment and evaluation contract exists for priority placement decisions, with explicit action feasibility, reward trade-offs, constraint signals, baselines, and known modeling assumptions.
• Candidate policies can be compared reproducibly on held-out workload and hardware conditions, with decision-quality, variance, drift, and failure-mode results that make limits visible rather than obscuring them in aggregate rewards.
• For experiments that progress beyond offline evaluation, evidence packages define predeclared success measures, bounded authority, stop/rollback conditions, and results from shadow or other appropriately controlled evaluation; Safety & Verification retains permissioning authority.

Where this role fits

Owns policy learning and evidence; independent controls decide whether an action is permitted.