Founding Kernel Engineer
Luminal
Founding Kernel Engineer
Luminal
Serverless AI deployment with GPU kernels
Hands-on experience developing GPU kernels.
Prior work on kernel autoresearch frameworks is a very strong positive.
Experience with ML compilers is a strong positive, but not required.
Hardware performance analysis, memory hierarchies, parallel execution, and benchmarking.
CUDA, Triton, GPU instruction sets, or compiler systems such as MLIR/LLVM and PyTorch backends.
Rust, open-source compiler contributions, or work on other accelerators.
No degree requirement. Researchers, recent PhDs, and candidates with substantial relevant projects are welcome.
Luminal (YC S25) builds an AI compiler and serving stack that optimizes models for faster, more efficient execution across accelerators. We're a small team working in person in San Francisco.
Research and build faster kernels for AI workloads. You'll work closely with the compiler team, investigate performance bottlenecks, and take optimizations through correctness checks, benchmarking, and execution on real hardware.
Our focus includes:
Agent-driven systems that write, test, and improve kernels.
Automated kernel research and search-based performance optimization.
More efficient execution for inference and LLM training.
Candidates who require a US visa are welcome to apply.
San Francisco, CA, US · Full-time · On-site. Annual base salary: $150,000–$250,000 USD. Equity: 0.15%–0.75%. Visa: Will sponsor.
Luminal provides a serverless cloud platform for deploying AI models with greater speed and reliability. Its main offering is an open-source machine learning compiler that automatically translates PyTorch models into optimized GPU kernels (e.g., Flash Attention) without manual tuning, working directly with CUDA and Metal through Rust. The platform offers features like auto-scaling, auto-batching, and on-demand weight streaming to handle varying traffic and minimize idle GPU time, acting as a drop-in upgrade to PyTorch to simplify production deployment. Differences from competitors include an end-to-end compiler-driven path that generates high-performance GPU code without extra engineering and a focus on compute-lean, low-overhead deployment in a serverless framework. Luminal aims to make it easier and cheaper to move AI models from research to production by reducing the engineering burden and speeding up inference, with YC-backed funding and a San Francisco base.
Explore more opportunities at Luminal