The RoleWe'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
