Member of Technical Staff, Measurement & Data Infrastructure

Member of Technical Staff, Measurement & Data Infrastructure

Make observations across unlike machines comparable, traceable, and useful for learning.

Status Open

Area Data Systems

General Diffusion is building compute world models across unlike machines; this role supplies the measurement layer that lets a workload run, environment, placement action, and observed outcome be compared with their full context intact. You will turn GD-X testbed and runtime observations into versioned, reproducible evidence for research and systems teams—without owning model training, placement policy, or the independent safety decision.

01 / The work

• Define and evolve versioned data contracts for workload shape, hardware and software environment, experiment or placement action, and observed outcome, including units, sampling semantics, and compatibility rules.
• Build ingestion and canonicalization paths that make measurements from heterogeneous machines comparable while retaining the raw records, calibration context, and derivation history needed to interpret them.
• Attach end-to-end provenance to datasets and derived slices: instrumentation and schema versions, workload revision, environment configuration, transformations, and quality status.
• Produce reproducible training and evaluation snapshots with explicit cohort definitions across architecture, workload, and time, so analyses can be rerun and compared without hidden data leakage.
• Implement automated checks for missing, late, duplicate, conflicting, or out-of-range records; surface schema breaks, instrumentation changes, and distribution shifts before they silently affect downstream conclusions.
• Publish clear producer and consumer interfaces with Runtime, Fleet, Research, and Safety teams, and make evidence fitness and known measurement limits visible rather than implicit.

• Experience building telemetry, event, or data infrastructure that serves multiple producers and consumers in a distributed systems environment.
• Strong schema and data-contract judgment, including backward-compatible evolution, stable identifiers, units, and semantics that remain interpretable as systems change.
• Hands-on practice with lineage, dataset versioning, manifests, and reproducible data or ML evaluation workflows.
• Measurement rigor: you can reason about sampling, clock alignment, aggregation, calibration, confounders, and the difference between an observation and an inference.
• Ability to implement data-quality validation and operational diagnostics using tools such as SQL and Python, then explain the resulting limits clearly to systems and research partners.

• A representative cross-architecture run can be traced from its raw records through its canonical dataset slice, with a versioned manifest that identifies the workload, environment, instrumentation, transformations, and applicable quality checks.
• Research and systems partners can independently reconstruct a selected training or evaluation cohort from a declared snapshot and obtain documented completeness, validity, and comparability results rather than relying on an informal export.
• The quality framework demonstrably catches seeded and observed failure modes—such as missing fields, incompatible units, schema changes, or conflicting measurements—and records the alert, triage context, and disposition.

Owns evidence infrastructure, not model training, policy learning, or safety sign-off.