Member of Technical Staff — Developer Technology

About The Role

RadixArk is seeking a Member of Technical Staff, Developer Technology (DevTech) to make LLM inference and training dramatically faster, cheaper, and more accessible on modern GPU hardware. Our systems sit at the center of how modern AI is served and trained: SGLang is a high-performance inference engine that serves trillions of tokens daily across leading AI companies and research labs, and Miles is our reinforcement-learning post-training framework for large-scale LLM and MoE models. Your work directly advances our mission to democratize AI: every improvement you ship lowers the cost and raises the ceiling of what developers everywhere can build.

As our technical face to a community of expert users and partners, you'll push the performance of SGLang and Miles through the lens of real production workloads. You'll profile and optimize GPU performance, enable new models and hardware, build kernels, deliver day-0 model support, and push the limits of inference and training. Working in close partnership with leading teams across the ecosystem, you'll turn their hardest, most ambiguous problems into concrete wins and clear guidance, and feed those improvements back into our systems and future roadmap.

Key Responsibilities

Accelerate AI workloads . Profile and optimize GPU performance for real production workloads on current and next-generation hardware, root-causing bottlenecks from kernels to distributed multi-node systems. Go deep in one or two focus areas.

The team collectively covers the full stack; each engineer specializes in one or two tracks: Inference performance : engine tuning, benchmarking, long-context and multi-turn optimization, parallelism strategy, production debugging Kernels and model/hardware enablement : custom CUDA/ROCm/Triton kernels, low-precision quantization, day-0 support for new models on new silicon Speculative decoding : draft-model training, acceptance-rate tuning, cross-platform kernel adaptation Training systems : RL post-training with Miles, FP8 training, elasticity, long-rollout and long-context efficiency Partner directly with the ecosystem. Turn ambiguous, high-stakes problems from expert engineers at our key partners into concrete wins, clear technical guidance, and reproducible cookbooks. Enhance SGLang and Miles .

Feed user-driven improvements back into our open-source systems and roadmap, so every win compounds across the ecosystem.

Qualifications

Minimum

Deep familiarity with LLM inference internals: distributed serving, parallelism, routing, KV-cache management, scheduling. Experience with low-precision quantization and inference/training (FP8, INT8/INT4; NVFP4 or MXFP4 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: pre-training, SFT, RL post-training, elasticity, 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.

Requirements

4+ years of experience in GPU systems, LLM infrastructure, or performance engineering. Strong profiling and debugging skills: able 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. Comfortable making progress on hard, ambiguous problems with little context to start from, and fast to ramp into unfamiliar systems, codebases, and domains. Ability to translate ambiguous asks into clear technical plans, verified cookbooks, and actionable recommendations, and to communicate credibly with expert engineering audiences.