ML Engineer

About the role

Kestrel trains perception models for warehouse robotics. Our research team is thirty people and our platform team is four, which is the problem you would be helping solve: training runs that are slower than they should be and harder to reproduce than they need to be.

What you will do

Own the distributed training platform — currently FSDP across sixty-four A100s
Cut time-to-first-batch for researchers, presently over forty minutes
Build experiment tracking and reproducibility that people actually use
Manage the data pipeline feeding four terabytes a day of sensor logs

What we look for

Multi-node distributed training in production
Real profiling ability — you have found and fixed a bottleneck people missed
Kubernetes and infrastructure-as-code
Patience for internal tooling, which is most of this job

What the assistant changed, and why

Change

Retitled to "ML Platform Engineer" from "Machine Learning Engineer"

Why

This is a platform role, and your last two years have been platform work that your title was hiding.

Change

Led with the FSDP migration instead of the model work

Why

It matches their exact stack, and it was buried under model results that this team is not hiring for.

Change

Added the time-to-first-batch improvement from your internal tooling work

Why

They name this metric specifically. You improved it by eleven minutes and had not mentioned it anywhere.

Change

Promoted Kubernetes and Terraform out of a trailing skills list

Why

Named requirements should not be in the last line of the resume.

What matched

FSDP multi-node training, exactly their stack
Profiling work with a measured throughput result
Kubernetes and Terraform in production
Two years of internal platform tooling

What it would not stretch

Role is on-site in Boston and you are listed as remote-preferred. Flagged before you approve — this one may not be worth applying to.

Awaiting your approval

The draft is written and the form is filled in on Kestrel Robotics’s own site. It sits here until you approve this one application. There is no approve-all, and nothing is submitted without you.