Research Engineer, ML Infrastructure

Plsn, Palo Alto, California

What you will do

Build and own distributed training systems across GPU clusters. Job launchers, checkpointing, recovery and monitoring should keep experiments moving when something fails.
Profile training end to end. Find bottlenecks in data loading, communication, memory and compute, then fix the ones that matter.
Build the tooling to launch, track and compare experiments. Make runs reproducible and shorten the path from an idea to a result.
Build reliable data pipelines for training and evaluation. Own data quality, throughput and the ability to reproduce a run.
Debug failures across GPUs, networking, numerics and data. Build systems that recover cleanly instead of leaving researchers to start over.
Implement and tune distributed parallelism where it helps. Understand the hardware and the tradeoffs rather than treating the training stack as a black box.

Plsn, Palo Alto, California

$150K–$220K

a year

This base pay is below the median for AI research jobs in the Bay Area with no level in the job title.

We compare the middle of each posted range. This is base pay as posted. The group does not include this job.

The Plsn site opens in a new tab.

Compare your résumé with this job. See where you match, what is missing and which lines to change.

Build and own distributed training systems across GPU clusters. Job launchers, checkpointing, recovery and monitoring should keep experiments moving when something fails.
Profile training end to end. Find bottlenecks in data loading, communication, memory and compute, then fix the ones that matter.
Build the tooling to launch, track and compare experiments. Make runs reproducible and shorten the path from an idea to a result.
Build reliable data pipelines for training and evaluation. Own data quality, throughput and the ability to reproduce a run.
Debug failures across GPUs, networking, numerics and data. Build systems that recover cleanly instead of leaving researchers to start over.
Implement and tune distributed parallelism where it helps. Understand the hardware and the tradeoffs rather than treating the training stack as a black box.