MTS & Research: GPU Kernel & Hardware RL Specialist

Collinear AI, Inc.; San Francisco (CA); On-site; USD 160,000 - 400,000; Full time

Job summary

Collinear in California is seeking aMTS - Research (Kernels & Hardware RL) to build the data engine for frontier AI. You will develop high-fidelity reinforcement learning environments and evaluation stacks that teach advanced agents to write efficient GPU kernels and solve hardware optimization problems.

You will own environments end to end, from selecting meaningful tasks and reference implementations to building reliable scoring and preventing reward hacking, collaborating daily with the

Qualifications

• Bachelor’s, Master’s, or PhD in CS, Electrical or Computer Engineering, or related field.
• Strong C/C++ and CUDA with production-quality Python; kernel development and memory hierarchy understanding.
• ML fluency with LLMs and evaluation knowledge.
• Experience conducting reproducible performance experiments and profiling.

Responsibilities

• Build agentic environments for GPU/accelerator kernels and hardware-aware optimization.
• Develop verification tests and performance graders; harden execution sandboxes against timing exploits and reward hacking.
• Analyze agent trajectories and failure modes; curate high-signal tasks and data; support post-training experiments across unseen workloads and hardware constraints.
• Collaborate daily with the founders and research staff to shape the roadmap and push the state-of-the-art in AI reliability.
• Scale infrastructure with reliable compilation, execution, profiling, and evaluation pipelines across hardware targets.

Skills

C/C++

CUDA

Python

Education

CS/EE/CE degree

Job description

Collinear in California is seeking aMTS - Research (Kernels & Hardware RL) to build the data engine for frontier AI. You will develop high-fidelity reinforcement learning environments and evaluation stacks that teach advanced agents to write efficient GPU kernels and solve hardware optimization problems.

You will own environments end to end, from selecting meaningful tasks and reference implementations to building reliable scoring and preventing reward hacking, collaborating daily with the