Staff ML Engineer – AWS Trainium & SageMaker

Robots & Pencils; United States; On-site; USD 140,000 - 200,000; Full time

Job summary

Robots & Pencils is seeking an engineer who can operate and train models on Amazon SageMaker with AWS Trainium, integrating from hardware-level concepts through PyTorch to production pipelines. You’ll work hands-on beyond APIs, shaping training requests at the accelerator and compiler level.

Strong PyTorch expertise, production experience with SageMaker, and comfort near the hardware layer are required. You’ll diagnose issues, optimize distributed training, and collaborate with client teams to

Qualifications

• Hands-on PyTorch experience with distributed or multi-device training.
• Production experience with Amazon SageMaker for training and inference.
• Comfort working close to accelerator hardware: understanding compilation and device behavior.

Responsibilities

• Train and operate models on Amazon SageMaker with AWS Trainium as the underlying compute.
• Write and optimize PyTorch training code with a real understanding of how it compiles and executes on Trainium (NeuronCore architecture, compiler behavior, memory and throughput tradeoffs).
• Diagnose training run issues that arise from hardware, not just the model, distinguishing data or code problems from compiler or device-level ones.
• Translate a request for "a Trainium job" into an actual working, cost-aware training pipeline, end to end.
• Tune distributed training runs for throughput and cost on SageMaker's training infrastructure.
• Work directly with client and internal engineering teams to scope and deliver real production training workloads.

Skills

PyTorch

Python

Distributed training

SageMaker

Tools

AWS Trainium

Neuron SDK

SageMaker

Job description

Robots & Pencils is an AWS Partner building production AI systems for enterprise clients who need real engineering, not proofs of concept that never ship. We work forward-deployed, embedded directly with client teams, solving the problems that are too new or too specialized for a typical vendor relationship to handle.

The Role

We're looking for an engineer who can operate and train models on Amazon SageMaker running on AWS Trainium, AWS's custom silicon built specifically for large-scale model training. This isn't a role where you call an API and wait. You'll be walking up the stack: understanding what a training request actually looks like at the Trainium hardware and compiler level, then carrying that understanding all the way up through PyTorch training code and into a production SageMaker pipeline.

PyTorch is the backbone of this work. If you know the framework deeply and you're comfortable reasoning about how your code actually behaves on custom accelerator hardware rather than treating it as a black box, this role is built around that skill set specifically.

What You'll Do

• Train and operate models on Amazon SageMaker with AWS Trainium as the underlying compute
• Write and optimize PyTorch training code with a real understanding of how it compiles and executes on Trainium (NeuronCore architecture, compiler behavior, memory and throughput tradeoffs)
• Diagnose training run issues that show up specifically because of the hardware, not just the model, distinguishing a data or code problem from a compiler or device-level one
• Translate a request for \"a Trainium job\" into an actual working, cost-aware training pipeline, end to end
• Tune distributed training runs for throughput and cost on SageMaker's training infrastructure
• Work directly with client and internal engineering teams to scope and deliver real production training workloads, not experiments that stay in a notebook

What You'll Bring

• Strong, hands-on PyTorch experience, ideally including distributed or multi-device training
• Production experience with Amazon SageMaker for training and/or inference
• Comfort working close to the hardware layer: you understand device-specific compilation and can debug issues that are actually about the accelerator, not just the model
• AWS Trainium or Inferentia (Neuron SDK) experience is a strong plus; if you don't have it yet but have deep PyTorch and a track record of picking up new hardware targets fast, we want to talk to you
• Solid Python fundamentals and comfort operating in a client-facing, production engineering environment