About the role
The Compute team makes sure the accelerators Clexerl owns do useful work. We write the kernels, compiler passes and low-level libraries that sit between model code and hardware, and we measure everything against the roofline: what the chip could do, what it does, and why the gap exists.
Training and inference both depend on this layer. A fused attention kernel or a better memory layout shows up as shorter training runs for Frontier AI researchers and lower serving cost for Clexerl Go agents. New model architectures often arrive with operations no vendor library handles well, and this team closes that gap.
We pair closely with researchers and the training and inference teams. Most work starts with a profile and a hypothesis, and ends with a benchmark that anyone can rerun.
What you’ll do
- Write high-performance kernels for attention, normalisation, mixture-of-experts routing and quantised matrix multiplication, targeting tensor cores and the memory hierarchy.
- Build and extend compiler passes that fuse operators, choose tilings and generate code for new accelerator generations.
- Profile training and inference workloads, attribute time to compute, memory and communication, and propose changes at the right layer.
- Validate numerics rigorously, including low-precision formats, so speedups never silently cost model quality.
- Maintain a benchmark suite and regression tracking so performance work stays visible and does not erode.
- Advise researchers on model shapes and sequence lengths that run efficiently on the hardware.
What you’ll bring
- Deep experience writing GPU kernels in CUDA, Triton, or a comparable accelerator programming model, with shipped results.
- Strong grasp of memory hierarchies, warp-level execution, bank conflicts, occupancy and asynchronous copy pipelines.
- Familiarity with an ML compiler stack such as MLIR-based or graph-level compilers, including how lowering and fusion decisions are made.
- Fluency in C++ and Python, and comfort reading assembly-level output when a kernel misbehaves.
- A habit of measuring first: roofline analysis, hardware counters and careful microbenchmarks over guesswork.
- Working knowledge of floating-point formats including bf16, fp8 and block-scaled variants.
Nice to have
- Experience with more than one accelerator architecture or vendor toolchain.
- Prior work on collective communication kernels or overlapping compute with interconnect traffic.
- Contributions to an open-source kernel library or compiler project.
- Experience with sparse or structured-sparse computation.
Work you might take on
- 01
Write a fused attention kernel for a new long-context variant that matches reference numerics and beats the library baseline on our shapes.
- 02
Add an fp8 grouped matrix multiplication path for mixture-of-experts layers, with per-block scaling and an automated accuracy check.
- 03
Extend the compiler to fuse normalisation, residual and quantisation steps into one pass and show the memory traffic saved.
- 04
Build a nightly roofline dashboard that tracks achieved utilisation per kernel across hardware generations and flags regressions.
Compensation
$320K – $540K
The US base salary range for this role. Every offer also includes equity in Clexerl and, when it spins out, in the venture you help build. We tell you the reasoning behind your number.
Benefits
Health, fully covered
Medical, dental and vision for you and your dependants, premiums paid in full.
Two kinds of equity
Equity in Clexerl, plus a stake in the venture you help build when it spins out.
22 weeks parental leave
Fully paid, for every parent, however your family arrives.
Rest that sticks
Flexible time off with a four-week minimum we actually check on, plus a company-wide week in December.
Learning and compute
A yearly budget for courses, conferences and books, and compute credits for side projects.
401(k) with a 4% match
Matched from your first paycheque and vested immediately.
How we hire
Apply 5 minutes
Send links to what you have built and a few lines on why this role. No cover letter needed.
Intro call 30 minutes
A conversation with the hiring manager about the work, the team and what you want next.
Work sample 2–4 hours
A paid take-home or a pairing session on a real-shaped problem. You pick the format.
Studio day One day
Meet the team, go deep on your work and ours. Remote or in person.
Offer Within 5 days
A written offer with the full range, the equity and the reasoning behind both.
Opens an email with a short template. Links beat résumés.