Member of Technical Staff, AI-Driven Compilation

About the role

SF Tensor is building the world’s fastest GPU compiler by using reinforcement‑learning agents that search a massive program space and prove correctness at the end. The Member of Technical Staff will own the search system that discovers optimal instruction selection, scheduling, tiling, fusion and phase ordering across diverse hardware targets.

Key responsibilities

• Design and implement RL agents that explore massive compilation search spaces and evaluate them on real silicon.
• Build agentic compilation loops that leverage LLMs to reason about IR, propose transformations, and learn from measured latency.
• Create reward‑assignment mechanisms based on wall‑clock performance and formal correctness proofs.
• Develop representations and embeddings for compiler IR that support learned optimization.
• Construct training infrastructure for distributed rollout and evaluation on GPUs, TPUs, AMD and Apple Silicon.

Required profile

• Strong background in reinforcement learning with hands‑on experience training agents.
• Experience building LLM‑based agents, tool use, or other agentic systems.
• Familiarity with GPU programming concepts and willingness to work at the ISA level.
• Proficiency in PyTorch or JAX.
• Ability to design and run rigorous experiments that produce trustworthy results.

Required skills

• Reinforcement Learning
• Large Language Model agents
• GPU programming
• PyTorch
• JAX
• LLVM (including custom backends)
• XLA, TVM, Triton, MLIR
• Program synthesis, formal methods, SMT solvers
• RLHF, reward modeling, preference learning
• GPU performance optimization, profiling, microbenchmarking

What we offer

• Base salary $275,000–$315,000 per year, plus meaningful equity.
• Comprehensive benefits and relocation assistance.
• Opportunity to ship learned kernels into production within days.
• Work onsite in our San Francisco office, collaborating with a small frontier‑scale team.