Sf Tensor Member of Technical Staff for AI-Driven Compilation to own the search process for GPU kernel optimization using agentic and RL systems. Responsible for building systems that decide which programs to try over a large search space.
Responsibilities
We build the fastest GPU compiler in the world. Most compilers have to preserve correctness at every transform, constraining how far they can search, while we prove correctness at the end instead, allowing us to search a far wider space, with agents, with RL, with anything that works and still guarantee the result. It's why we hold #1 on NVIDIA's own kernel benchmark across hundreds of production kernels.
We're hiring a Member of Technical Staff for AI-Driven Compilation to own the search itself. Proving correctness at the end is what makes the search legal and your job is to make it good.
You'll be building the agentic and RL systems that decide which programs are worth trying (instruction selection, scheduling, barriers and stall counts, tiling, fusion, phase ordering) over a space that is enormous precisely because nothing in it has to be conservative.
The reward signal here is unusually clean for RL: measure wall-clock on real silicon gated by a formal proof. No proxy metrics, no reward hacking that survives contact with the verifier. The system already does things worth seeing. Given only a naive attention spec, the search found FlashAttention v4-level performance on a B300 in about 25 minutes, but it also generalizes. The same loop runs on AMD, TPU and Trainium as well as targets we'd never seen before, such as Apple Silicon.
You'll have access to better tooling than anywhere else because we own the stack all the way down to the ISA, allowing us to create kernels that others can't even express. For example, our custom LLVM backend emits cubins directly, without ptxas, letting us work on instruction selection, scheduling, register allocation and stall counts. Our team understands the hardware better than anyone else, to this extent we've built a bit-exact software model of Blackwell's tcgen05.
The kernels you create will ship immediately on runs such as pre-training AlphaFold v3 at 3.4× the throughput, post-training robotics models on Trainium or running our custom RL rollout engine on TPU at multi-100B parameter scale.
You'll design and implement RL systems that search over a massive program space: instruction selection, schedules, tile sizes, fusion strategies and phase ordering
You'll build agentic compilation loops that use LLMs to reason about IR, propose transformations and learn from measured results
You'll design the search and credit-assignment machinery around an exact reward built from measured latency on real hardware and a formal correctness proof
You'll create representations and embeddings of compiler IR that hold up under learned optimization
You'll build the training infrastructure for compiler optimization agents, including rollout throughput and distributed evaluation on real silicon
You'll push transfer and cold-start performance so that the search works on unfamiliar targets and new ISAs from the first trial
Qualification
Someone proficient in PyTorch or JAXSomeone familiar with RLHF
Required
Someone with a strong background in reinforcement learning with hands-on experience training agents
Someone with experience building LLM agents, tool use and other agentic systems
Someone with familiarity in GPU programming concepts and a willingness to get down to the ISA
Someone proficient in PyTorch or JAX
Someone with the ability to design and run experiments that produce trustworthy results
Preferred
Someone with experience in ML compiler stacks (XLA, TVM, Triton, MLIR) or LLVM backends
Someone with a background in program synthesis, superoptimization, combinatorial search, formal methods, SMT solvers or verified compilation
Someone familiar with RLHF, reward modeling or preference learning
Someone with research contributions in RL or learned optimization
Someone familiar with GPU performance optimization, profiling and microbenchmarking
