KERNEL ENGINEER

MakerMaker; San Francisco (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

MakerMaker in San Francisco hires a kernel engineer to write and optimize GPU kernels and supporting systems software that accelerates training and inference workloads.

You will work closely with researchers to close the gap between functional correctness and hardware efficiency, ensuring end-to-end performance improvements across the stack.

Qualifications

• 4+ years writing performant GPU kernels and related systems.
• Fluency with memory hierarchy, occupancy, and tensor cores.
• Experience profiling with Nsight, ncu, or similar tools.
• Track record of kernel-level optimizations in real systems.
• Strong systems understanding of kernel placement within frameworks.
• Comfort reading Python and C++ around kernels.

Responsibilities

• Write and optimize GPU kernels for training and inference workloads.
• Profile workloads using hardware counters and translate to kernel optimizations.
• Co-design kernels with research teams when needed.
• Integrate optimized kernels into training and serving stacks and benchmark.
• Maintain kernel quality as hardware and frameworks evolve.
• Spread kernel fluency across the team.

Skills

GPU kernel development

Profiling skills

Systems design

C++ / Python proficiency

End-to-end optimization

Tools

Nsight

ncu

CUDA

ROCm

Job description

ABOUT THE COMPANY

We're building autonomous research agents for recursive self-improvement (multi-agent systems that propose, run, and analyze machine learning experiments). We're a small, well-funded team that works onsite in San Francisco.

ABOUT THE ROLE

You'll write and optimize the GPU kernels and supporting systems software that makes our training and inference workloads fast. This is deep, low-level work (performance counters, memory bandwidth, warp-level scheduling) applied to the specific shapes and patterns our models actually use.

We hire kernel engineers because the gap between \"this works\" and \"this is fast on the hardware we have\" is enormous, and that gap directly bounds what our researchers can try. You'll close that gap.

WHAT YOU'LL DO

• Write and optimize GPU kernels (CUDA, ROCm, Triton, or similar) for training and inference workloads: attention variants, MoE layers, custom activations, communication primitives
• Profile real workloads with hardware counters and translate findings into specific kernel-level optimizations
• Co-design kernels with the research teams, when the kernel and the algorithm need to change together, you participate in both
• Integrate optimized kernels into our training and serving stacks; benchmark before and after; verify the win is real end-to-end
• Maintain kernel quality over time as hardware, frameworks, and workloads shift underneath
• Spread kernel-level fluency across the team; we want this expertise shared, not siloed

WHAT WE'RE LOOKING FOR

• 4+ years writing performant GPU kernels (CUDA, ROCm, Triton, or production-grade equivalent)
• Hardware-level fluency: memory hierarchy, occupancy, register pressure, tensor cores, warp scheduling
• Profiling fluency (Nsight, ncu, or comparable tools) and the discipline to measure before changing
• Track record of shipping kernel-level optimizations that moved a measurable metric in a real system
• Strong systems expertise: you understand how kernels live inside larger frameworks and how integration choices affect end-to-end performance
• Comfortable reading framework-level Python and C++ around your kernels

NICE TO HAVE

• Open-source contributions to kernel libraries, compilers, or ML frameworks
• Experience with multiple accelerator architectures (different GPU families, TPUs, custom ASICs), preferably AMD GPUs
• Familiarity with collective communication primitives (NCCL or equivalent)
• Compiler or runtime background

THIS ROLE IS PROBABLY NOT FOR YOU IF

• You haven't gotten your hands dirty at the kernel level: this isn't a higher-level systems role rebranded
• You want to stay narrowly in one library; we expect breadth across the kernel surface our models actually use
• Performance work without measurable end-to-end impact frustrates you