Staff Engineer, Local Inference & Kernel Optimization

Sonder; New York (NY); On-site; USD 250,000 - 300,000; Full time

Benefits offered by this job

Health, dental, and vision coverage

Flexible PTO

Relocation support

Job summary

Sonder is seeking an inference and performance engineer to own the systems layer between our models and the hardware they run on—starting with Apple Silicon and macOS. You will make our models faster, smaller, and more power-efficient through kernel optimization, memory layouts, and quantization.

You will work closely with researchers to co-design inference systems, profiling from workloads to measured improvements, and shipping reliable local inference stacks across the Mac platform.

Qualifications

• Deep experience in ML inference, GPU programming, or numerical computing with shipped improvements.
• Strong C++ skills and hands-on kernel writing in Metal, CUDA, Triton, or similar environments.
• Understanding GPU architectures and memory hierarchies, bandwidth, and synchronization impacts on kernels.
• Experience optimizing matrix multiplication, attention, or similar heavy ops, with numerical validation.

Responsibilities

• Own end-to-end performance: profile latency, throughput, memory, and energy across workloads.
• Build kernels where it matters: optimize core ops with tiling and memory layouts to reduce traffic.
• Make low-precision inference useful: evaluate quantization and mixed precision for performance and quality.
• Improve the runtime: reduce overhead and enable reuse of computation across screen content changes.
• Co-design for consumer hardware: align architecture decisions with Mac memory and compute.
• Build performance infrastructure: reproducible benchmarks and checks across Apple Silicon generations.

Skills

ML inference

GPU programming

C++

Kernel optimization

Profiling

Quantization

Model optimization

System-level thinking

Tools

Metal

CUDA

Triton

Job description

Sonder is seeking an inference and performance engineer to own the systems layer between our models and the hardware they run on—starting with Apple Silicon and macOS. You will make our models faster, smaller, and more power-efficient through kernel optimization, memory layouts, and quantization.

You will work closely with researchers to co-design inference systems, profiling from workloads to measured improvements, and shipping reliable local inference stacks across the Mac platform.