Senior Inference Optimization ML Engineer (AI)

Location: Mountain View, United States

Company

hirify.global is building generalist intelligent robots with an end-to-end stack spanning high-performance hardware, robot systems, infrastructure, foundation world models, and manufacturing.

What you will do

• Own end-to-end inference optimization for vision, video, language, proprioception, and action foundation models, improving latency, throughput, and cost per inference.
• Port and optimize models across AMD GPUs, Google TPUs, wafer-scale and dataflow accelerators, consumer Blackwell GPUs, and Jetson/Thor-class robot hardware.
• Deploy and optimize on-robot inference under latency, power, thermal, and memory constraints, including decisions about robot-cloud workload splitting.
• Build cost/performance models and hardware recommendations for serving, offline evaluation, and on-robot use.
• Develop portable and vendor-specific kernels, optimize attention, KV cache, memory layout, and fused operators, and apply quantization, pruning, distillation, and model or tensor parallelism.
• Build cross-hardware benchmarking, regression, numerical-parity, policy-quality, and robot-behavior validation infrastructure while collaborating with research, robotics, and infrastructure engineers.

Requirements

• 2+ years of experience in inference optimization, ML systems, ML compilers, or a closely related field.
• Deep hands-on experience with PyTorch.
• Strong understanding of compute, memory bandwidth, memory capacity, and interconnect bottlenecks across hardware architectures.
• Experience with quantization, including INT8, FP8, FP4, and AWQ; distillation, pruning, and compilation using tools such as torch.compile, XLA, TensorRT, IREE/MLIR, TVM, or vendor compilers.
• Familiarity with inference serving frameworks and their hardware support, including vLLM, SGLang, JetStream, and Triton Inference Server.
• Excellent debugging and measurement skills, with the ability to identify bottlenecks, design experiments, and validate improvements.
• A high-ownership, pragmatic approach focused on total solution cost rather than peak benchmark results.

Nice to have

• Inference optimization for diffusion, multimodal, or video models, including variable-length sequences and packing or bucketing.
• Production or deep project experience with accelerator stacks beyond NVIDIA datacenter GPUs, or serious consumer-GPU deployment experience.
• Kernel authoring across multiple backends, including HIP, Composable Kernel, Pallas, Triton, CUTLASS, or CuTe.
• Contributions to open-source ML compilers or hardware backends such as PyTorch, OpenXLA, Triton, vLLM, IREE, or TVM.
• Experience with wafer-scale systems, TPU pods, consumer-GPU fleets, Jetson/Thor platforms, or TensorRT, JetPack, DLA, power, and thermal profiling.
• Experience with speculative decoding, continuous batching, disaggregated prefill/decode, streaming, low-latency robot control, or working with hardware vendors on driver, compiler, or early-access silicon issues.

Culture & Benefits

• Direct ownership of optimization work that affects inference cost, model evaluation, and robot deployment scale.
• Benchmarks and hardware recommendations directly influence accelerator purchasing and platform strategy.
• High ownership within a small, specialized team responsible for the optimization layer across all target hardware.