Our client builds expert-written benchmark tasks that train and evaluate AI agents on real command-line work. They're looking for GPU kernel and ML performance engineers to design expert-level Terminal-Bench tasks: real problems from professional kernel work, turned into end-to-end challenges an AI agent has to solve on its own.

Each task you build includes a realistic workplace, a working reference solution, and an automated check that objectively proves whether the agent got it right. If you know how a fast FP8 kernel can quietly return the wrong answer, and how you would catch it, this work was made for you.

This is a task-based compensation model: you are paid per completed task, not per hour. Details are covered in the application questions.

What You'll Do

- Design expert-level Terminal-Bench tasks drawn from real GPU kernel and ML performance work, framed as challenges an AI agent must solve end to end
- Define the files, systems, and environment each task runs on, so the problem is realistic, reproducible, and reflects production tools and data
- Write a complete, working reference solution for every task (for example, a fast FP8 CUDA kernel) and verify it runs correctly in the environment
- Build automated tests that objectively prove correctness, catching silent failures like numerical drift,



precision loss, and metric bugs
- Calibrate difficulty so a typical task takes an expert about 60–75 minutes to solve: hard for today's AI agents, but fair and unambiguous
- Document each task clearly (instructions, expected outputs, and how mistakes would be detected) and revise based on reviewer feedback

What You Bring

- 3+ years of hands-on experience writing and tuning GPU kernels at real scale (beyond notebooks or API calls), or a PhD with active research in the field
- Hands-on CUDA or Triton experience and GPU profiling, with recent work (last 3 years) where you can name the systems, tools, data, and constraints involved
- The ability to explain how ML results go wrong without raising errors (numerical drift, metric bugs) and how you'd catch it
- Fluency in Python or another language and the command line, plus working with files, CSV/JSON, APIs, or databases
- Availability to take on tasks on a flexible basis starting early November 2026

Nice to Have

- AI lab or LLM infrastructure experience
- Contributions to PyTorch, JAX, or Megatron
- NCCL and multi-GPU cluster performance experience
- Docker and Git proficiency
- Experience writing automated tests

📌 GPU Kernel Engineer: Design Expert Benchmark Tasks for AI Agents (India)
🏢 Toptal
📍 India