Senior System Software Engineer, Agentic Inference - Dynamo

NVIDIA•Us, CA

15d•Hybrid

About The Position

We are looking for a Senior System Software Engineer to work on Dynamo. NVIDIA is hiring software engineers for its GPU-accelerated deep learning software team. Academic and commercial groups around the world are using GPUs to power a revolution in AI, enabling breakthroughs in problems from image classification to speech recognition to natural language processing. We are a fast-paced team building Generative AI inference platform to make design and deployment of new AI models easier and accessible to all users.

Requirements

Masters or PhD or equivalent experience
10+ years in Computer Science, Computer Engineering, or related field
Ability to work in a fast-paced, agile team environment
Excellent Rust/Python programming and software design skills, including debugging, performance analysis, and test design.
Understanding of modern LLM API semantics, including structured outputs, tool calling, reasoning controls, token accounting, context management, and multimodal inputs.

Nice To Haves

Prior contributions to open-source AI inference frameworks (e.g., vLLM, TensorRT-LLM, SGLang).
Experience optimizing GPU memory, KV and prefix caches, or high-performance networking for long-context, reasoning, and tool-calling workloads.
Understanding of LLM-specific inference challenges for agentic workloads, including context and reasoning-token growth, bursty tool-call-driven traffic, multi-turn state reuse, and scheduling across concurrent trajectories.
Prior experience integrating self-hosted LLM serving stacks with agent harnesses such as OpenCode, Codex, Claude Code, and Pi, including compatibility for APIs, streaming, structured outputs, tool calls, and session semantics.

Responsibilities

Develop open source software to serve inference of trained AI models running on GPUs.
Contribute to the development of disaggregated serving for Dynamo-supported inference engines (vLLM, SGLang, TRT-LLM) and expand these capabilities to support agentic inference workloads, including long-horizon reasoning, tool calling, and stateful, multi-turn execution.
Innovate in inference-state management for long-running agents, including KV- and prefix-cache reuse and transfer across heterogeneous memory and storage hierarchies with NIXL, to reduce repeated prompt processing, improve latency and token throughput, maximize GPU utilization, and lower per-token and per-task costs for self-hosted LLMs.
Build and evolve Dynamo’s distributed inference frontend across vLLM, SGLang, and TensorRT-LLM, delivering day-0 support for new models, model-specific request parameters, upstream API compatibility, and stateful Responses API semantics.
Balance a variety of objectives: build robust, scalable, high performance software components to support our distributed inference workloads; work with team leads to prioritize features and capabilities; load-balance asynchronous requests across available resources; optimize throughput under latency constraints; and integrate the latest open source technology.