Member of Technical Staff, Heterogeneous Runtime & Placement

What you’ll work on

• Define the execution contract that maps workload requirements, target capabilities, runtime compatibility, and resource availability into explicit placement eligibility checks before any ranking occurs.
• Build the runtime and control-plane paths that bind approved workloads to supported targets, reconcile intended versus observed execution state, and preserve a safe fallback when a target cannot execute the work.
• Make each placement decision inspectable: capture the eligibility result, selected target, relevant runtime and workload context, fallback or re-placement action, and execution outcome.
• Implement bounded re-placement and reversal paths for failures, incompatibilities, or degraded execution; exercise those paths through canaries and recovery drills rather than relying on an untested rollback claim.
• Develop conformance and lifecycle tests across supported runtimes and target classes, including admission, launch, recovery, and cleanup behavior under partial failure.
• Partner with compiler, GPU systems, fleet, measurement, and safety teams to consume capability, cost, topology, and guardrail signals at clear interfaces; do not absorb their code-generation, physical-fleet, or authorization ownership.

What you bring

• Hands-on experience shipping or operating a scheduler, inference-serving system, distributed runtime, or control plane where a decision must be carried through to real execution.
• Practical judgment about heterogeneous workload lifecycles: device and runtime compatibility, resource admission, dependency or environment mismatches, partial failure, and recovery.
• Ability to turn ambiguous placement constraints into explicit contracts, test cases, and observable state transitions rather than opaque heuristics.
• Strong systems-debugging and telemetry skills, including tracing a decision from requested resources through binding, execution, and fallback.
• Experience designing guarded rollout, canary, and rollback mechanisms with an operational understanding of idempotency, reservations, and failure containment.

What progress looks like

• A reviewed execution contract and automated conformance matrix demonstrate that known-incompatible target/runtime combinations are rejected before ranking, while valid alternatives retain explicit eligibility evidence.
• Shadow or canary execution records provide an end-to-end trail for representative supported workloads—from request and placement rationale through observed binding and outcome—with failure and fallback states distinguishable in the data.
• A documented reversal drill demonstrates re-placement or rollback for selected compatibility and runtime-failure scenarios, with prior state, decision history, and the independent safety boundary preserved for audit.

Where this role fits

Owns production decision execution, not model training or kernel authoring.