New opportunity

Machine Learning / Reinforcement Learning Infrastructure Engineer

Eka Robotics · Cambridge, United States

About this role

Eka Robotics is seeking a hands-on engineer to build the training infrastructure for robot learning. The role focuses on making it easier for researchers to test ideas, reproduce experiments and scale successful work from prototypes to large-scale distributed training.

Responsibilities

Design, implement and maintain model-training systems, including job orchestration, scheduling, checkpointing and experiment tracking.
Build intuitive tools for launching, monitoring, debugging and reproducing experiments, with an emphasis on researcher productivity.
Work with researchers to scale reinforcement learning and machine learning pipelines across compute clusters, manage cloud compute resources efficiently and contribute to core JAX model and training code.
Develop infrastructure that supports research methods while guiding best practices for reliable training at scale.

Minimum qualifications:

A BS, MS or higher degree in Computer Science, Computer Engineering, Machine Learning or a related technical field.
Strong software engineering fundamentals and a track record of building ML training infrastructure, internal developer platforms or scalable systems.
Hands-on large-scale training experience with JAX, PyTorch or TensorFlow; JAX is preferred.
Familiarity with distributed training, multi-host setups, data pipelines, and cloud platforms or orchestration systems such as Kubernetes, SLURM, GCP or AWS.
Experience building automated testing pipelines, CI/CD for ML workflows and custom logging or telemetry stacks. Strong cross-functional communication and ownership are also required.

Preferred qualifications include experience in robotics, reinforcement learning or other machine learning systems, and designing abstractions that balance researcher flexibility with system reliability. This is a full-time, on-site position listed in the Boston Area; the posting's structured location identifies Cambridge, Massachusetts.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Machine learning infrastructureReinforcement learningDistributed trainingJAXPyTorchTensorFlowKubernetesSLURMGoogle Cloud PlatformAWSJob orchestrationSchedulingCheckpointingExperiment trackingData pipelinesCloud computingCI/CDAutomated testingLoggingTelemetrySoftware engineeringSystems designCross-functional communication

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Own reliable training systems
The role owns large-scale training infrastructure, including orchestration, scheduling, checkpointing and experiment tracking. Prepare to explain how you would design and maintain these as a coherent system.
Reduce friction for researchers
The posting emphasizes intuitive tools for launching, monitoring, debugging and reproducing experiments. Prepare to discuss how tooling can speed research without obscuring what a training run did.
Scale distributed ML and RL training
Working with researchers to move pipelines from prototypes to reliable runs across compute clusters is central to the role. Prepare to discuss multi-host training and data-pipeline considerations.
Use cloud compute efficiently
Efficient allocation and utilization of cloud-based compute are explicit responsibilities. Prepare to explain how you would assess utilization and make scheduling decisions as demand grows.
Bridge research code and infrastructure
The engineer will partner with researchers, guide training practices and contribute to core JAX model and training code. The posting also calls for ML testing, CI/CD and telemetry experience; prepare to connect these practices to researcher workflows.

Likely challenges to think through

Flexibility versus reliable abstractions
A likely design tension, not a stated problem: novel research methods may need freedom that standardized tooling restricts. Prepare to describe where you would offer extension points and where you would enforce consistent run metadata and safeguards.
Prototype speed versus distributed reproducibility
A possible difficulty is preserving debuggability when an experiment moves from a quick prototype to a multi-host run. Prepare to reason through configuration capture, checkpoint recovery and ways to compare runs across scales.
Compute efficiency versus operational resilience
A likely trade-off is that more checkpoints and telemetry can aid recovery and diagnosis but consume resources. Prepare to explain how you would evaluate those costs alongside utilization and failed-run recovery.

Practice your answers

Technical practice

How would you design a training platform that takes experiments from prototypes to distributed runs?

What to coverThe role owns large-scale training systems, including orchestration and scheduling. Preparation suggestion: Explain how you would structure job submission, resource requests and failure handling while keeping small experiments easy to launch.

How would you make training checkpoints reliable and useful for recovery?

What to coverCheckpointing is an explicit infrastructure responsibility. Preparation suggestion: Discuss what state must be saved, how to validate a checkpoint and the trade-off between checkpoint frequency and training overhead.

What would you record to make an experiment reproducible?

What to coverThe posting emphasizes experiment tracking and tools for reproducing runs. Preparation suggestion: Address code, configuration, data and checkpoint provenance, and explain how you would detect when two runs are not meaningfully comparable.

How would you diagnose poor scaling in a multi-host training run?

What to coverScaling ML and reinforcement learning pipelines across compute clusters is a stated duty. Preparation suggestion: Describe how you would distinguish compute, communication and input-pipeline bottlenecks before changing the training setup.

How would you contribute to core JAX training code without compromising infrastructure reliability?

What to coverThe role includes contributing to JAX model and training code; JAX is preferred, while PyTorch and TensorFlow are also accepted experience. Preparation suggestion: Discuss how you would reason about training correctness, performance measurements and regression tests for a code change.

How would you keep a data pipeline from limiting training throughput?

What to coverThe qualifications call for familiarity with data pipelines and large-scale training. Preparation suggestion: Explain what you would measure to separate data loading delays from model-compute limits, and how you would evaluate a proposed pipeline change.

How would you allocate cloud compute efficiently across training workloads?

What to coverEfficient allocation and utilization of cloud-based compute resources are explicit responsibilities. Preparation suggestion: Compare scheduling choices using utilization, wait time and the effect of interruptions on long-running experiments.

What testing and CI/CD checks would you put around ML training workflows?

What to coverThe posting asks for experience with automated testing pipelines and CI/CD for ML workflows. Preparation suggestion: Distinguish fast code checks from more expensive training checks, and explain how you would catch failures before a large run.

What telemetry would help researchers debug a failing or slow experiment?

What to coverThe role calls for monitoring and debugging tools plus custom logging and telemetry stacks. Preparation suggestion: Describe useful job, training and resource signals, and how you would connect them to a specific run without overwhelming users with noise.

How would you design a launch interface that is flexible for researchers but reliable at scale?

What to coverIntuitive experiment tooling and the balance between researcher flexibility and system reliability are both stated priorities. Preparation suggestion: Explain which choices you would expose, which safeguards you would standardize and how you would evaluate whether the interface reduces friction.

Behavioural practice

Tell me about a time you translated a researcher's needs into an infrastructure change.

What to coverThe role requires close collaboration with researchers. Preparation suggestion: Use a real situation, the questions you asked and actions you took, then describe the result for the research workflow.

Describe a time you took ownership of a training-system failure.

What to coverThe posting emphasizes ownership and reliable training at scale. Preparation suggestion: Explain the situation, how you investigated and communicated it, the fix you made and the observed result.

Tell me about a time you balanced user flexibility against system reliability.

What to coverDesigning abstractions with that balance is a preferred qualification. Preparation suggestion: Give a real example of the competing needs, your decision and its result, including what you would change in hindsight.

Describe a time you improved a developer tool based on user feedback.

What to coverImproving researcher productivity through intuitive tooling is central to the role. Preparation suggestion: Cover the initial friction, how you gathered feedback, what you changed and the resulting effect on users.

Sharpen your knowledge

Topics to prepare

Training-system architecture
Posting: The role owns large-scale training infrastructure from prototypes through distributed runs. Revision exercise: sketch the components and failure boundaries for launching, scheduling, tracking and resuming a training job; explain which responsibilities belong in the training code versus the infrastructure.
Distributed training and data flow
Posting: Multi-host training, data pipelines and scaling ML and reinforcement-learning workloads across clusters are explicit. Revision exercise: diagram how data, model state and progress move through a multi-host run, then explain what changes when a prototype scales out.
Scheduling and compute utilization
Posting: The engineer manages cloud compute allocation and utilization. Revision exercise: compare scheduling choices for several competing training jobs, including how to avoid idle resources and handle a job that needs more capacity.
Experiment reliability and reproducibility
Posting: Checkpointing, experiment tracking and reproducible runs are central responsibilities. Revision exercise: define what a run must record to be repeatable and reason through recovery after an interrupted run.
Researcher-facing abstractions
Posting: Tooling should make experiments easy to launch, monitor and debug while balancing flexibility with reliability; that balance is also a preferred qualification. Revision exercise: critique a proposed experiment-launch interface from both a researcher's and an infrastructure owner's perspective.

Skills to revise

Work through a JAX training change
Posting: The role contributes to core JAX model and training code; large-scale experience with JAX, PyTorch or TensorFlow is required, with JAX preferred. Suggested exercise: trace a small JAX training step, make one controlled change and explain how you would verify its behavior in a larger run.
Specify an orchestrated training job
Posting: Job orchestration, scheduling and cluster workloads are explicit; Kubernetes and SLURM are examples, not mandatory tools. Suggested exercise: draft a job specification and launch-to-completion workflow in one of those systems, including resource requests and failure handling.
Implement checkpoint and run tracking
Posting: Checkpointing, experiment tracking and reproduction tooling are responsibilities. Suggested exercise: build a small training run that saves and restores state, records its configuration and results, and demonstrates a repeat run.
Test and observe an ML workflow
Posting: Automated testing pipelines, CI/CD for ML workflows, and custom logging or telemetry stacks are minimum-qualification experience areas. Suggested exercise: add automated checks to a training workflow and instrument a failed run so its cause can be investigated.
Translate a research need into tooling
Posting: The engineer partners with researchers and is expected to communicate across functions. Suggested exercise: rehearse clarifying an experiment-launch or debugging request, then propose a small interface and explain its reliability trade-offs. This is interview preparation, not a stated employer exercise.

Your preparation checklist

Prepare a training-infrastructure project walkthrough
The role calls for ownership of large-scale training systems. If you have a relevant project, prepare a diagram and a concise account of what you built, the decisions you owned, failures you addressed, and measurable results.
Practice a distributed-training design exercise
Sketch a proposed path from prototype to multi-host training, covering job orchestration, scheduling, checkpoint recovery, data pipelines, experiment tracking, and efficient cloud-compute use. Label it as your design, not Eka Robotics’ existing architecture.
Gather framework and scaling evidence
Prepare a concrete example of large-scale training with JAX, PyTorch, or TensorFlow, including debugging and scaling tradeoffs. JAX is preferred in the posting; discuss core JAX training code only if you have relevant experience.
Demonstrate a researcher-facing workflow
The posting emphasizes tools for launching, monitoring, debugging, and reproducing experiments. Draft a short workflow showing how a researcher would run and reproduce a training experiment, and explain where flexibility must be balanced with reliability.
Prepare reliability and observability examples
Select examples, if available, of automated tests, CI/CD for ML workflows, and custom logging or telemetry. Be ready to explain what each detected, how failures were investigated, and what changed afterward.
Rehearse a collaboration and ownership story
The role involves partnering with researchers and guiding training practices. Prepare a specific account of how you translated a research need into an infrastructure decision, handled disagreement, and followed through.
Write questions for the interviewers
Ask how researchers currently move from prototypes to distributed runs, where experiment reproducibility or compute utilization causes friction, and which infrastructure improvements would matter most first.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…