Member of Technical Staff - ML Ops

Every physical thing you touch exists because somebody successfully navigated the laws of physics: the chips in your phone, the vehicles on the road, the data centers powering AI. Physics determines what can be built, how well it performs, and where it breaks. Yet the tools engineers use to understand physical behavior are too slow and too specialized to use continuously while designing, so critical decisions get made with only a partial view of how a system will behave.

Our mission is to make physical reasoning as accessible to engineers as language became through modern AI. This is not an attempt to build slightly better engineering software. It is an attempt to change how physical products are designed. Our technology is used today by many of the world's most advanced semiconductor and electronics organizations, including nearly half of the twenty largest companies in the industry.

We are backed by Khosla Ventures and Eclipse Ventures.

Our foundation model learns physics from simulations, and we generate the simulation data ourselves. The simulations encapsulate real-world, validated physical results. Our aim is to create a data corpus covering multiple types of physics over an exhaustive set of physical objects.

Our data corpus is expanding. We currently operate on the petabyte scale, and our data volume is growing rapidly. We are hiring the engineer who makes that possible: the person who owns the pipeline from "here is a generation recipe" to "here is a versioned, validated dataset a training run can consume," and the ML lifecycle around it. Your job is to expand our job scheduling and execution platform to ensure that campaigns of millions of simulations, ML training runs, and future jobs are simple to run, fault-tolerant, and well-cataloged for traceability.

This is hands-on engineering in a production codebase. You will write, test and ship the code yourself.

• Scale the data generation pipeline by orders of magnitude. Our generation framework runs real simulations on 3D parts, and we are scaling from thousands of simulations to millions.

• Own dataset lineage and reproducibility. Every completed simulation records how it was produced and the recipe how it can be augmented for future data generation jobs.

• Build data quality into the pipeline. Acceptance criteria, bad-sample flagging, coverage and distribution checks against the target parameter space, and per-field statistics on every sample.

• Run the job platform. Our GPU scheduler uses Temporal as a durable control plane and Ray as the execution plane, with priority, admission, cancellation and per-job isolated environments. You will extend it as generation, training and validation compete for the same fleet.

• Close the loop between data and models. You will own training job launch (Ray Train, torchrun), checkpoint management and the model registry, and the path that puts a trained checkpoint back into the pipeline, for example as a warm start for the next generation campaign or as an inference pipeline in production.

• Integrate with the infrastructure team. Our compute is managed by our infrastructure experts, and you will work closely with them to ensure compute and data resources will meet workload needs.

• Make costs visible. Know what a sample costs to generate, store and serve, by physics type and by configuration. Find the stage that wastes GPU time, and make the case for which data is worth generating next.

• Keep the codebase healthy as it grows. Write tests first, keep CI fast, write schema migrations that are safe on populated databases, and document what you build so the next engineer can operate it without you.

• 5+ years building production data or ML pipelines, including at least 2 years where you owned a pipeline that ran on GPUs at meaningful scale

• Experience running large batch or simulation workloads: thousands of long-running jobs, partial failures, retries, preemption, and making sure work is neither lost nor duplicated

• Strong Python and software engineering fundamentals: clean abstractions, typed code, and a test suite you trust

• Comfort with relationalrelationa and No
  
  SQLl databases at scale: schema design, migrations on live data, bulk writes, and query performance once tables reach hundreds of millions of rows

• Experience with workflow orchestration or distributed execution frameworks, such as Temporal, Ray, Airflow, Argo, Kubernetes Jobs or Slurm

• Working knowledge of cloud object storage and of the file…