Member of Technical Staff, ML Engineer

Physical Superintelligence is a startup with roots at Google, NVIDIA, Harvard, Meta, MIT, Oxford, Johns Hopkins, Cambridge, and the Perimeter Institute building AI systems to discover new physics at scale. We are seeking an ML engineer to be the engineering arm of our researchers and teams: turn their code into systems that run at scale, and build and run the training and inference infrastructure those systems depend on.

Our mission is to discover and commercialize transformative physics breakthroughs at scale with artificial superintelligence, safely, verifiably, and for broad public benefit.

The last century's golden age of physics gave us transistors, lasers, and nuclear energy. We believe artificial superintelligence will unlock the next one. We're creating the infrastructure to industrialize scientific discovery and usher in this new era.

• →

  Help researchers and the other teams with the engineering: turn research code into systems other people can run, with contracts, tests and an on-call owner.

• →

  Build the services researchers use to launch runs, route across models and ship results, and be the first call when one breaks.

• →

  Own the training and inference infrastructure those services run on: distributed jobs, GPU scheduling, model serving, on the shared platform Engineering runs.

• →

  Go deep in one lane: distributed training, inference serving, the ML platform, or the gateway and agent infrastructure.

• Three or more years building and operating ML training or inference infrastructure in production, at a company that trains or serves models at meaningful scale. You have owned a system through a failure.

• Three or more years of software engineering on systems others depend on. You can build a service that other engineers and researchers depend on every day, with a contract, tests that catch a wrong answer, and a consumer who is not you.

• Depth in at least one of the following. We hire for one lane and expect you to learn the others on our platform:

  • Distributed training: PyTorch DDP and FSDP, Ray or comparable, multi-node jobs; checkpointing and resume; where step time goes.

  • LLM inference serving: vLLM, SGLang, Triton or comparable; batching, KV cache, quantization; an endpoint behind a gateway.

  • ML platform and MLOps: Kubernetes GPU scheduling and preemption, experiment tracking, model and checkpoint registries, metering.

  • LLM gateway, agent runtime and evaluation infrastructure: API proxying and provider routing with capture and metering; agent frameworks, sandboxes, memory and task registries; eval harnesses that run and grade agents at scale.

• Enough ML fluency to debug a training loop, a reward signal or inference-time behavior alongside a researcher.

• A second lane from the list above.

• GPU infrastructure, CUDA, or performance engineering for ML workloads.

• Cloud infrastructure (GCP, AWS) and infrastructure as code.

• Prior work embedded with a research team, turning research code into production systems.

We hold a high technical bar and give people full ownership of their work, from spec to ship to on-call. We write contracts before logic, test against real systems instead of mocks, and favor simple designs that ship over clever ones that do not. Our development process is AI-native: we work with agentic coding tools daily, write specs that are legible to humans and agents alike, and lead with leverage.