Member of Mathematical Staff

SF Tensor Company; San Francisco (CA); On-site; USD 225,000 - 315,000; Full time

Benefits offered by this job

Relocation assistance

Job summary

SF Tensor is building the fastest GPU compiler and seeks a Member of Mathematical Staff to own the math behind the proof engine. You will work with a team on formal models for new hardware, numeric stability, and proofs across diverse architectures, guiding research agendas and collaborating with engineers on verification tasks.

We offer access to advanced tooling, collaboration with researchers, and relocation assistance for SF-based work in San Francisco.

Qualifications

• PhD in Math is strongly recommended but not required
• Good judgement about when a full proof is worth it and when a cheaper guarantee is enough
• Pragmatic about tooling; know when SMT works, when an interactive proof is necessary and when custom decision procedures are necessary
• Comfortable owning an open-ended problem where you set the research agenda yourself
• Ability to explain a proof strategy to an engineer who doesn't know the math, and push back when corners are cut
• Contributed to Mathlib, Flocq or seL4
• Hands-on experience with Lean or Rocq
• Work or publications in the field of program verification
• Work or publications in numerical analysis or numerics
• Work or publications in optimization theory or dynamics of stochastic training
• Deep knowledge in floating-point formalization

Responsibilities

• Own the proof engine end-to-end, ensuring we can prove kernels across all edge cases
• Collaborate with the engineering team on building formal models of new hardware
• Formalize numeric stability and model convergence across diverse architectures
• Research around predicting training divergence during runs
• Collaborate with the research team on post-training proofs and program verification
• Formalize low-precision computation and its numerical stability
• Collaborate with the engineering team on formally proving sandboxes safe

Education

PhD in Math

Tools

Lean

Coq

Job description

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 guarantees the result. It's why we hold #1 on NVIDIA's own kernel benchmarkacross hundreds of production kernels.

Our core proof machinery is real and it works, but right now there are some rough edges that limit the number of kernels we can prove.With the rise of AI capabilities in mathematics combined with our bit-identical models and tools, such as our own model for large-scaleLean proof writing, things that were nearly impossible just a few months ago are now totally viable, but we need someone who figures outwhat theory we're missing, builds it and then coordinates with the engineering team to deploy it at scale, which is why we are hiringa Member of Mathematical Staff to own the math behind the proof engine and other surfaces for applying rigorous math to an increasinglywishy-washy domain.

This is not an advisory role and you'll have access to better tooling than anywhere else, such as bit-exact software models of hardware componentslike the tcgen05 , a compiler capable of emitting code at a lower level than publicly possible, andfrontier models with capabilities tailored to what your work needs, trained with our team.

What You'll Do

• You'll own the proof engine end-to-end, ensuring we can prove kernels across all edge cases
• You'll collaborate with the engineering team on building formal models of new hardware
• You'll work on formalizing numeric stability and what drives model convergence across diverse architectures
• You'll work on research around the extent to which we can predict training divergence during runs
• You'll collaborate with the research team on our post-training efforts around formal proofs and program verification
• You'll work on formalizing low-precision computation and its numerical stability
• You'll collaborate with the engineering team on formally proving sandboxes safe

What We're Looking For

• A PhD in Math is strongly recommended but not required: what matters is what you've proven and built
• Someone with good judgement about when a full proof is worth it and when a cheaper guarantee is enough
• Someone who's pragmatic about tooling and knows when SMT works, when an interactive proof is necessary and when custom decision procedures are necessary
• Someone who's comfortable owning an open-ended problem where you set the research agenda yourself
• Someone who can explain a proof strategy to an engineer who doesn't know the math, and push back when they want to cut corners

Nice to Have

• Someone who's contributed to projects such as Mathlib, Flocq or seL4.
• Someone with hands-on experience working with Lean or Rocq
• Someone who's done work or published in the field of program verification
• Someone who's done work or published in the field of numerical analysis or numerics
• Someone who's done work or published in the field of optimization theory or the dynamics of stochastic training
• Someone with deep knowledge in the field of floating-point formalization

Why Join Us

You will sit in the only place in the world right now working on guaranteeing the correctness and training stability of frontier models, which is one of the most important problems of our time, withthe tools and freedom to actually solve it. The truth is that most formal methods work never leaves the paper and most ML work never gets near a proof. Here, the proof engine decides what is allowed toship into production, so your work sits on the critical path of everything we do and gets tested against more kernels in a day than most verification tools see in their lifetime.

We invest heavily in AI tooling for mathematics, from our own model for large-scale Lean proof writing to bit-exact models of the hardware, so you spend your time on the theory and not the tedium, and get from idea proof as fast as possible. Beyond that, the people who reverse-engineered the tensor cores, write the compiler and train the models you are proving and using for your mathematics are the same people you eat lunch with, so when a proof hinges on what the silicon actually does, the answer is a conversation away.

We're a small team operating at frontier scale. We pre-trained foundation models on 4,000 AMD GPUs as a team of three, designed andbrought up GB300 NVL72 clusters and designed a TOP500 supercomputer.

We believe that hard problems get solved in person and most of our work happens at our office in San Francisco. We offer relocationassistance and, where possible, we'd like you here as often as possible.

The base salary range for this full-time position is $225,000-$315,000, plus meaningful equity and benefits.