Machine learning jobs at US tech companies

Hot sub-areas in 2026

Foundation-model post-training (RLHF, DPO, RLAIF) is the most in-demand specialty — almost every model lab is hiring. Inference-side work (CUDA kernels, custom serving stacks, KV-cache tricks, quantization) is the second hottest band. Multi-modal and agent-runtime roles are growing fast.

Classical ML roles (recsys, ranking, fraud, search) are still abundant at FAANG-tier companies and high-growth marketplaces. If you don't want to compete in the foundation-model arms race, filter on those titles instead — the bar is more accessible and the work is often higher impact per dollar.

What employers actually look for

Across the listings on this page, the most-mentioned skills are PyTorch, distributed training (FSDP / DeepSpeed / Megatron), Kubernetes, and Python. CUDA / Triton appear on roughly a third of MLE roles. Strong systems fundamentals (latency budgets, batching, memory) show up more often than novel research output.

Publication record matters most for Research Scientist titles. For ML Engineer and Applied Scientist roles, demonstrated production impact (eval harnesses you built, models you shipped, throughput you optimized) tends to weigh more than papers. Lumi's resume tailor maps your background to each listing's specific requirements.

How to apply efficiently in a hot market

ML hiring moves fast. The strong roles get 200-400 applicants in the first 48 hours; recruiters skim, not read. Sign up for Lumi and we'll match each new role against your resume, score the fit, and email you within an hour of any high-match listing going live. That's the difference between landing in the first stack of resumes and the third.