About Netpreme

The biggest bottleneck in AI systems today is memory. Chips compute faster than they can be fed data—a problem known as the memory wall—and Netpreme is building the solution that allows AI memory to scale independently of AI compute. Our approach spans the full stack, from a proprietary memory management framework at the XPU kernel and OS levels to X-Mem™, the world’s first Memory Processing Unit (MPU) silicon. Founded by MIT researchers at the intersection of optical networking, ML systems, and GPU kernel optimization, we’re based in Cambridge, MA, and Santa Clara, CA. We’re looking for people who want to own hard problems end to end—and help shape the next generation of AI systems.

Click here to read our latest blog!

About the Role

Serving frontier LLMs efficiently at scale is a moving target: every new model and workload reshuffles the tradeoffs across parallelism, batching, KV caching, and scheduling. You’ll own Netpreme’s inference-serving stack end to end, from tuning vLLM, SGLang, and TensorRT-LLM to architecting distributed serving with NVIDIA Dynamo (prefill/decode disaggregation, KV-cache-aware routing). Working directly with our hardware and systems engineers, you’ll build the benchmarks and drive the optimizations that take new models from research to production.

This role will be performed onsite from one of our offices in Santa Clara, CA or Boston, MA (relocation assistance available).

The Opportunity

Impact: AI is advancing faster than the infrastructure underneath it. We’re working at that layer, unlocking more capability from every GPU while dramatically improving efficiency. The work you do here compounds across state-of-the-art models, systems, and real-world applications.

Timing: Joining now means real ownership, including equity, and a direct say in product direction. You’ll work from first principles, move quickly from insight to execution, and see your work show up in what we ship, not just in a roadmap.

Culture: We care about rigor, clarity, and impact. Disagreement is welcome when it’s thoughtful, the best idea wins regardless of who it comes from, and you’ll learn something new most weeks.

Your Impact

Deploy and operate models in production: Run large language and multimodal models using inference engines like vLLM, SGLang, and TensorRT-LLM.

Own the serving infrastructure: Build and maintain distributed inference systems with NVIDIA Dynamo, Kubernetes, Helm, and Ansible, including upgrades, monitoring, troubleshooting, and reliability.

Optimize performance: Tune parallelism strategies (TP/EP/DP/PP), batching, KV caching, routing, and scheduling to meet latency, throughput, and SLA targets.

Benchmark and measure: Build reproducible benchmarks for latency (TTFT, TPOT, E2E), throughput, concurrency scaling, and system efficiency.

Support diverse workloads: Enable tool calling, agentic and coding workflows, RAG, and long-context applications.

Bring new tech into production: Evaluate new models, frameworks, and optimizations, and partner with hardware, model, and systems engineers to ship improvements.

Qualifications

BS, MS, or PhD in Computer Science, Computer Engineering, or a related field, or equivalent practical experience.

2+ years of engineering experience in LLM inference, ML systems, GPU systems, or related infrastructure.

Hands-on experience deploying and optimizing at least one LLM inference engine (e.g., vLLM, SGLang, or TensorRT-LLM).

Solid understanding of LLM inference fundamentals: prefill/decode, batching, KV caching, and latency/throughput tradeoffs.

Strong Python and Linux skills, plus hands-on experience with Kubernetes and containerized deployments.

Comfortable benchmarking and troubleshooting systems, and communicating clearly in a small engineering team.

Nice to Have

Experience with Helm, Ansible, or similar deployment and configuration tools.

Experience with distributed inference systems like NVIDIA Dynamo, including parallelism strategies (TP/EP/DP/PP), prefill/decode disaggregation, and KV-cache-aware routing.