ML Research Engineer (Performance Engineering)

Metamorphic•Palo Alto, CA

About The Position

Metamorphic is developing new approaches to intelligence by combining machine learning with large-scale experimental neuroscience, informed by the principles that make the brain efficient, flexible, and robust. We are building foundation models trained on rich, continuous neural data — a high-resolution model of the brain at a scale never before possible. Our founding team spans machine learning, neuroscience, and neurotechnology, with prior work including the MICrONS project, Neuropixels, and the Enigma project, as well as foundational scientific contributions in learning, neural computation, and generative modeling. Our work sits at the frontier of AI research, and we believe the highest-impact discoveries will come from researchers and engineers working as a single, tightly collaborative team. The name Metamorphic reflects our belief that the next advances in intelligence will come from a change in form, beyond scale — from artificial to natural intelligence. About the Role We are seeking Research Engineers to join our growing AI research team. You will be responsible for maximizing the training and inference performance of Metamorphic's foundation models, from quantization and low-precision training, to MoE routing optimization, to writing custom CUDA/Triton kernels for our novel architecture. This is a high-impact, technically deep role at the frontier of ML research and engineering. You will write and optimize GPU kernels, profile and eliminate performance bottlenecks, tune low-precision training strategies, and work closely with researchers to ensure architectural decisions translate to efficient and scalable implementations. You'll have substantial autonomy to shape foundational technical decisions on a small, high-impact team.

Requirements

Bachelor's degree or higher in Computer Science, Machine Learning, or a related field
Strong software engineering skills with a proven track record of building complex systems
Strong proficiency in CUDA, Triton, or similar, with demonstrated experience writing and optimizing GPU kernels
Hands-on experience with mixed-precision and low-precision training and a practical understanding of numerical stability tradeoffs
Deep knowledge of transformer architectures at the implementation level
Experience with MoE architectures: routing algorithms, load balancing, and the systems-level challenges of expert dispatch across GPUs
Hands-on experience with GPU profiling tools (Nsight Compute, Nsight Systems, PyTorch Profiler)
Experience integrating, customizing, and extending third-party high-performance libraries (FlashAttention, cuDNN, Triton, Quack, or similar) into production training stacks

Nice To Haves

Experience with CUTLASS, cuDNN APIs, and NCCL internals
Familiarity with inference optimization techniques and serving frameworks
Familiarity with diffusion models or multimodal model architectures
Experience with inference optimization techniques (KV-cache management, speculative decoding, post-training quantization) and serving frameworks (vLLM, TensorRT-LLM)

Responsibilities

Maximizing the training and inference performance of Metamorphic's foundation models
Writing and optimizing GPU kernels
Profiling and eliminating performance bottlenecks
Tuning low-precision training strategies
Working closely with researchers to ensure architectural decisions translate to efficient and scalable implementations
Shaping foundational technical decisions on a small, high-impact team