Member of Technical Staff - ML Systems

• Technical depth in ML infrastructure, with experience in training systems, model serving, or distributed data and rollout pipelines. You can explain the architectural decisions and measurable improvements in work you've owned.
• Knowledge of how PyTorch, JAX, or similar frameworks operate beneath their APIs, with the ability to follow execution through the code, profile workloads, and diagnose correctness or performance issues.
• Strong distributed systems fundamentals, including resource scheduling, communication, storage, and recovery. You can reason about how these interact in demanding ML workloads.
• Understanding of the execution demands of post-training: how sampling, reward evaluation, optimization, and checkpoint updates fit together, and where efficiency or reliability can break down.
• Ability to evaluate systems changes through reproducible measurements and work with researchers and customers to decide which problems matter most.

Experience with distributed training frameworks, vLLM or SGLang, GPU profiling, container orchestration, or research platforms is particularly relevant. CUDA, Triton, and compiler or runtime expertise are useful specializations. We value depth in one area and the ability to work across the surrounding stack; kernel expertise is not required.

We welcome engineers from ML infrastructure teams, research groups, and startups, including earlier-career engineers with a substantial systems project. The depth of your contribution matters more than a particular title or tenure.

Develop the infrastructure for post-training experiments, from distributed rollout execution to training and checkpoint evaluation. Improve resource allocation, parallel execution, and recovery so researchers can run more experiments and spend less time managing them.

Make model serving efficient for both agent workloads and training data generation. Investigate bottlenecks in batching, KV-cache use, scheduling, and memory movement, and evaluate changes against latency, throughput, cost, and model quality. Depending on your specialty, this can include kernel and runtime optimization with CUDA, Triton, or related tools.

Build scalable systems for collecting trajectories, coordinating teacher models, and running agent environments. Make data generation and evaluation reproducible, with reliable execution and clear connections between datasets and the experiments that use them.

Develop benchmarks, observability, and automated experiment workflows that help researchers test ideas and understand results. Work with researchers and customers to turn new methods into dependable training and deployment capabilities, and share useful systems work through technical writing and open-source contributions where appropriate.

Tell us about a system you're proud of: what it needed to do, what you personally built, and a difficult reliability or performance problem you solved. Include a resume and any links to code or technical writing you'd like us to read. If your interests span systems and research, apply once and tell us where you'd most like to contribute.

Explore our published research https://nexttoken.co/research for examples of the work these systems support.

NextToken builds specialized models and the systems that help them improve. We work with AI product teams to turn production traces, expert feedback, and real tasks into evaluations, training data, and more capable models. Our goal is to make it practical for companies to develop intelligence of their own: models that understand their work and improve as that work evolves.

Our founding team has built search infrastructure and post-trained models at frontier AI labs and on large-scale product teams. You’ll work directly with the founders on research and engineering, help shape our technical direction, and see your work deployed in customer products.