Why This Matters at Sonder

About the role

Sonder is an applied AI lab building models that learn how people work. Today, AI largely depends on people explaining what they are doing and remembering when to ask for help. We are building private models that live on your computer, understand work as it happens, and learn the patterns, preferences, and judgment behind how you operate. Our first product, Twin, brings this intelligence to the Mac. It builds a memory of your work and uses that context to offer help at the right moment.

The Role: We are looking for an inference and performance engineer to own the systems layer between our models and the hardware they run on. You will make our models faster, smaller, and more power-efficient, working from model execution and quantization down to memory layouts and custom GPU kernels. We are starting with Apple Silicon and macOS, using MLX and Metal. Our models run throughout the working day, sharing memory and compute with the applications someone is using. Sustained power draw, memory bandwidth, and responsiveness under contention matter as much as peak throughput. A faster kernel matters when it makes the whole system better. You will work closely with researchers to co-design models and inference systems around these constraints, with ownership from profiling an unfamiliar workload to shipping a measured improvement.

Why This Matters at Sonder: Local intelligence has to earn its place on someone's computer. Every reduction in memory use or sustained power draw makes room for more capable models and longer context while keeping the Mac responsive. This work determines how much intelligence can stay private on the device, and how naturally Twin can fit into a person's working day.

Location: This role is based in New York. We work together in person and are building the early team here. Visa Sponsorship: We sponsor visas. We cannot guarantee success in every case, but if you are the right fit, we are committed to working through the process with you.

Responsibilities

• Own end-to-end performance. Profile latency, throughput, peak memory, and energy use across representative workloads. Separate time spent in vision encoding, prefill, and decoding, and identify whether the limiting factor is compute, memory bandwidth, or runtime overhead.
• Build kernels where they matter. Optimize critical operations such as matrix multiplication, attention, and quantized linear layers. Use tiling, fusion, and appropriate tensor layouts to reduce memory traffic while managing register pressure, occupancy, and synchronization.
• Make low-precision inference useful. Evaluate weight and KV-cache quantization, mixed precision, and dequantization costs against both performance and model quality. Work with researchers to understand which layers and workloads tolerate lower precision.
• Improve the runtime. Reduce allocation and kernel-launch overhead, manage caches, and schedule latency-sensitive inference alongside background work. Find opportunities to reuse computation as screen content and context change, and validate when reuse remains correct.
• Co-design for consumer hardware. Help