Build the agents that make inference cheaper.
We are a small early-stage team. The work is kernels, schedulers, quantization recipes, and the agents that search across them. Remote-first, with equity.
Three roles open now.
Kernel engineer
Write and profile attention, MoE dispatch and low-bit GEMM kernels in CUDA, HIP and Triton. Turn hand-found wins into templates the kernel agents can generate and verify on their own. You have shipped a kernel that beat the vendor library and can explain why.
Inference systems engineer
Live inside vLLM, SGLang and TensorRT-LLM. Own scheduling, KV-cache tiering, speculative decoding and prefill/decode disaggregation in the search space, and the correctness gate that guards them. You have run an LLM serving stack in production and been paged for it.
Founding go-to-market
First commercial hire. Run the design-partner program, own the pipeline, write the case studies with the team, and turn the first three partners into the first three customers. You have sold infrastructure to engineers and can read a latency chart.
Small, measured, in the open.
Numbers over narratives
Every change is a benchmark with a methodology. Internal reviews look like our research posts, because they become them.
Agents do the searching
People decide what to optimize and what correct means. Agents run the search. If you find yourself doing a grid search by hand, that is a feature request.
Early stage, honestly
We are pre-revenue and pre-GPU-fleet, running in customers’ clouds. Compensation is salary plus meaningful equity. We will tell you exactly where the company stands before you decide.
How to apply
Email hello@deepinference.co with the role in the subject line, a few sentences on the fastest thing you have made, and a link to it if it is public. A kernel, a serving patch, a benchmark write-up, a deal you closed. No cover letter needed. We reply to every application within a week.
Not a fit for these roles?
Tell us what you would build here anyway. The best roles we have opened started as an email.