Share of this job’s tasks that AI is already used for at work, from real usage. It is not a forecast of job loss, and a low number means no use seen yet. Source: Anthropic Economic Index.

Check how this applies to your own job

Skills this job asks for

Say which you have and which you want to build. We use it to match you to jobs like this one.

About the job

A NauWork client is seeking a Founding Engineer – Kernel & Compiler Lead to join their team. The position is on-site or hybrid in Sunnyvale, CA; remote employment may also be possible .

This early-stage startup is tackling one of AI’s toughest challenges: delivering faster intelligent compute while dramatically reducing energy consumption. Its hardware-software co-design approach uses probability-native silicon to rethink the deterministic architectures behind today’s AI infrastructure. The deeply technical founding team is building a next-generation compute platform designed to deliver major gains in performance and energy efficiency across modern AI workloads. An early hire will have an exceptional opportunity to influence the company’s technology, direction, and culture from the ground up.

This hands-on founding engineer will build and own the probabilistic kernel library and PyTorch backend that map modern generative models onto novel silicon. The role will lower the Core ATen operator set produced by the PyTorch 2.x compiler stack onto native hardware primitives, then extend coverage to world models and video generation models. Working directly with hardware architects, this engineer will help shape the primitive set and ISA while defining how kernels should operate on a new computing architecture. As the software stack and organization mature, they will also hire and lead the engineers who expand it.

Responsibilities:

Map the Core ATen ops set emitted by AOTAutograd and PrimTorch decompositions onto processor primitives—establishing which operators lower to native probabilistic execution, which are decomposed, and which need bespoke kernels.

Design, implement, and tune the probabilistic kernel library itself: sampling primitives, stochastic arithmetic, low-precision and stochastic-rounding paths, and the probabilistic reformulations of dense linear algebra and attention that make the architecture win.

Stand up the PyTorch backend end to end—dispatch key registration, custom TorchInductor/AOTAutograd compiler backend, operator registration, and the fallback paths that keep real models running while coverage is still growing.

Extend coverage to what world models and video generation models require.

Participate in architecture reviews and shape the primitive set and ISA the processor exposes. When a kernel is awkward to express, this engineer will be in the room where that gets fixed in silicon rather than worked around in software.

Required Experience:

Substantial hands-on experience building kernels, compilers, or runtimes for a non-CPU target—GPU, TPU, NPU, FPGA, or a custom AI accelerator.

Deep working knowledge of the PyTorch 2.x compiler stack: torch.compile, TorchDynamo, AOTAutograd, PrimTorch decompositions, the Core ATen opset, and TorchInductor.

Strong systems programming in C++ and Python, with real fluency in performance work—profiling, roofline reasoning, memory hierarchy, and data movement as the usual bottleneck.

Experience taking a kernel library from nothing to production coverage on new hardware, including the judgment calls about what to implement first.

Strong written and verbal communication. This engineer will be the interface between the hardware architects and everyone building models on top of the stack.

Experience with probabilistic or sampling-based hardware is not expected—almost no one has it. What matters is having brought up a software stack on unfamiliar silicon before and being genuinely energized, rather than deterred, by an architecture that does not behave like a GPU.

Bonus Experience:

Experience writing Triton, CUDA, or comparable low-level kernels, or building a custom TorchInductor backend.

Background in computer architecture or hardware/software co-design, particularly on accelerators the candidate helped define rather than inherited.

Hands-on work with diffusion models, video generation, or world models.

Familiarity with MLIR, OpenXLA/StableHLO, PJRT, ExecuTorch, or similar compiler infrastructure.

Experience with quantization, sparsity, mixed-precision inference, or other model-efficiency techniques at the kernel level.

To Learn More:

503-388-9585

833-NAU-WORK

careers.nauwork.com

Category: Technology – System Software

Position: Founding Engineer – Kernel & Compiler Lead

Location: San Francisco Bay Area – On-Site or Hybrid; Remote May Be Possible

Job Type: Direct-Hire, Full-Time