About The Position

RadixArk is seeking a Member of Technical Staff on its Developer Technology (DevTech) team to make LLM inference and training faster, cheaper, and more accessible on modern GPU hardware. The role advances SGLang, a high-performance inference engine, and Miles, a reinforcement-learning post-training framework for large-scale LLM and MoE models. Improvements contribute to production AI workloads and RadixArk’s mission to democratize AI infrastructure. As a technical point of contact for expert users and partners, you will optimize SGLang and Miles around real production workloads. The work includes profiling GPU performance, enabling new models and hardware, building kernels, delivering day-zero model support, and improving inference and training systems. You will work with ecosystem partners to solve complex technical problems, provide guidance, and incorporate user-driven improvements into RadixArk’s open-source systems and roadmap. Engineers specialize in one or two areas while collaborating across the broader stack.

Requirements

  • At least 4 years of experience in GPU systems, LLM infrastructure, or performance engineering.
  • Strong profiling and debugging skills, including the ability to root-cause performance and correctness issues across the stack.
  • Hands-on GPU programming experience in at least one of CUDA, ROCm, or Triton, and willingness to work across platforms.
  • Strong programming skills in Python plus C++ or CUDA.
  • Ability to make progress on difficult, ambiguous problems with little initial context and to ramp quickly into unfamiliar systems, codebases, and domains.
  • Ability to turn ambiguous requests into clear technical plans, verified cookbooks, and actionable recommendations, and to communicate credibly with expert engineering audiences.

Nice To Haves

  • Deep familiarity with LLM inference internals, including distributed serving, parallelism, routing, KV-cache management, and scheduling.
  • Experience with low-precision quantization and inference or training, such as FP8 or INT8/INT4; NVFP4 or MXFP4 experience is a strong plus.
  • Experience writing and optimizing custom GPU kernels.
  • Practical familiarity with speculative decoding methods such as Eagle, DFlash, or DSpark.
  • Working knowledge of large-scale distributed training, including pre-training, SFT, RL post-training, elasticity, and long-context workloads.
  • Experience optimizing across both NVIDIA and AMD platforms.
  • Hands-on experience with SGLang, Miles, vLLM, TensorRT-LLM, Megatron, or comparable frameworks.
  • Contributions to open-source AI/ML projects.

Responsibilities

  • Profile and optimize GPU performance for production workloads on current and next-generation hardware, identifying bottlenecks from kernels through distributed multi-node systems.
  • Specialize in one or two areas, which may include inference performance, kernels and model or hardware enablement, speculative decoding, or training systems.
  • Tune inference engines, run benchmarks, optimize long-context and multi-turn workloads, develop parallelism strategies, and debug production systems.
  • Build or optimize custom CUDA, ROCm, or Triton kernels; work on low-precision quantization and day-zero support for new models on new silicon.
  • Work on speculative decoding, including draft-model training, acceptance-rate tuning, and cross-platform kernel adaptation.
  • Improve training systems, including RL post-training with Miles, FP8 training, elasticity, and long-rollout and long-context efficiency.
  • Partner with expert engineers at key ecosystem partners to turn ambiguous, high-stakes problems into concrete improvements, technical guidance, and reproducible cookbooks.
  • Feed user-driven improvements back into SGLang, Miles, and the product roadmap.

Benefits

  • Equity
© 2026 Teal Labs, Inc
Privacy PolicyTerms of Service