GPU / CUDA Engineer

About Bespoke Labs

Bespoke Labs is a VC-backed applied AI research startup in Mountain View, CA, building core infrastructure and RL environments to train and evaluate intelligent agents. Home of OpenThoughts (100K+ monthly downloads, 200+ models trained) and Terminal Bench, a leading agentic coding benchmark used by frontier labs. Founded by ex-Google DeepMind and UC Berkeley faculty, advised by Jeff Dean.

The Role

You'll write, optimize, and debug CUDA kernels that directly power our AI systems — from training and inference to RL workloads. You'll also build the tooling (profilers, inference engines, benchmarks) that keep our systems at peak performance.

What You'll

• DoWrite and optimize CUDA kernels for GEMM, attention, MoE, and graph operations
• Use PTX assembly and architecture-specific techniques for Hopper/Blackwell hardware
• Apply memory coalescing, warp-level programming, tensor cores, and compute/memory overlap
• Integrate kernels into PyTorch, vLLM, Megatron, and TorchTitan
• Profile and debug with Nsight Systems, Nsight Compute, and Torch Profiler
• Build internal tooling and contribute to open-source GPU libraries

What We Need Must Have

Hands-on CUDA kernel optimization experience (kernel hacking strongly preferred)

• Strong grasp of GPU architecture — memory hierarchy, warp execution, synchronization
• Proficient in C/C++ for high-performance systems
• Experience in profiling and resolving GPU bottlenecks

Nice to Have

• Flash Attention or Transformer kernel optimization
• Cutlass, Triton, Thrust, or CUB experience
• Distributed/multi-GPU (NVLink, NCCL) background
• Open-source GPU contributions or published research

Why Join

High ownership. Frontier research. Real production impact. A small, elite team is building the infrastructure that the next generation of AI runs on