Responsibilities:
Design a unified AI Infra & Serving architecture platform for composite AI workloads such as LLM Training & Inference, RLHF, Agent, and Multimodal processing. This platform will integrate inference, orchestration, and state management, defining the technical evolution path for Serverless AI + Agentic Serving
Design a heterogeneous execution framework across CPU/GPU/NPU for agent memory, tool invocation, and long-running multi-turn conversations and tasks. Build an efficient memory/KV-cache/vector store/logging and state-management subsystem to support agent retrieval, planning, and persistent memory.
Build a high-performance Runtime/Framework that defines the next-generation Serverless AI foundation through elastic scaling, cold start optimization, batch processing, function-based inference, request orchestration, dynamic decoupled deployment, and other features to support performance scenarios such as multiple models, multi-tenancy, and high concurrency.
Key Requirements:
Strong foundational knowledge in system architecture, or computer architecture, operating systems, and runtime environments;
Hands-on experience with Serverless architectures and cloud-native optimization technologies such as containers, Kubernetes, service orchestration, and autoscaling
vLLM, SGLang, Ray Serve, etc.); understand common optimization concepts such as continuous batching, KV-Cache reuse, parallelism, and compression/quantization/distillation
Proficient in using Profiling/Tracing tools; experienced in analyzing and optimizing system-level bottlenecks regarding GPU utilization, memory/bandwidth, Interconnect Fabric, and network/storage paths
Proficient in at least one system-level language (e.g., C/C++, Go, Rust) and one scripting language (e.g., Python)
TPBN1_UKTJ