Lead cutting‑edge research on on‑device model and agent efficiency, focusing on compression, acceleration, and heterogeneous inference across CPU/GPU/NPU platforms.
Responsibilities
Conduct research on model compression and acceleration techniques such as FP4/INT4 QAT, speculative decoding, and on‑policy distillation
Develop algorithms for agent and token routing/device‑cloud orchestration
Design low‑latency, privacy‑preserving multimodal model structures and inference mechanisms
Define new research problems, design experiments, build prototypes, and evaluate benchmarks
Analyze algorithm‑system co‑design space for heterogeneous computing and OS constraints
Requirements
Master's degree in Artificial Intelligence, Computer Science, Electronic Information Technology, or related field
Design, build, and operate distributed training, rollout, and orchestration systems for large-scale LLM and multimodal post-training across multi-GPU, multi-node environments.
Compensation
Depending on background, skills, and experience, the expected annual salary range for this position is $200,000 to $400,000, plus equity.
Responsibilities
Design, build, and operate distributed training, rollout, and orchestration systems for large-scale LLM and multimodal post-training across multi-GPU, multi-node environments.
Profile and optimize performance across the full-stack — model implementation, parallelism strategies, communication libraries, and GPU kernels — to improve throughput, latency, memory efficiency, hardware utilization, and cost.
Investigate numerical correctness and low-precision issues in distributed training and inference, including train–inference consistency for reinforcement learning.
Improve the reliability of long-running workloads through checkpointing, fault recovery, observability, and operational tooling.
Build supporting infrastructure for reinforcement learning and agentic post-training, including asynchronous rollout, trajectory collection, sandboxed execution, evaluation harnesses, and data pipelines.
Requirements
Nice to have: RL algorithms such as PPO, GRPO
Nice to have: modern post-training frameworks such as Miles, slime, AReaL, verl, Prime-RL
Nice to have: key open-source contributions to training or inference frameworks such as SGLang, vLLM, Megatron-LM
Nice to have: GPU kernel development such as CUDA, Triton, CUTLASS
Nice to have: communication-layer optimization such as NCCL, RDMA, NVLink/NVSwitch