Senior Backend Engineer, Inference Platform
Together AI · San Francisco
About the Role Together AI is building the Inference Platform that brings the most advanced generative AI models to the world. Our platform powers multi-tenant serverless workloads and dedicated endpoints, enabling developers, enterprises, and researchers to harness the latest LLMs, multimodal models, image, audio, video, and speech models at scale. If you get a thrill from optimizing latency down to the last millisecond, this is your playground. You’ll work hands-on with tens of thousands of GPUs (H100s, H200s, GB200s, and beyond), figuring out how to fully utilize every FLOP and every gigabyte of memory. You’ll collaborate directly with research teams to bring frontier models into production, making breakthroughs usable in the real world. Our team also works closely with the open source community, contributing to and leveraging projects like SGLang, vLLM, and NVIDIA Dynamo to push the boundaries of inference performance and efficiency. Shape the core inference backbone that powers Together AI’s frontier models. Solve performance-critical challenges in global request routing, load balancing, and large-scale resource allocation. Work with state-of-the-art accelerators (H100s, H200s, GB200s) at global scale. Partner with world-class researchers to bring new model architectures into production. Collaborate with and contribute to the open source community, shaping the tools that advance the industry. A culture of deep technical ownership and high impact — where your work makes models faster, cheaper, and more accessible. Competitive compensation, equity, and benefits. Responsibilities Build and optimize global and local request routing, ensuring low-latency load balancing across data centers and model engine pods. Develop auto-scaling systems to dynamically allocate resources and meet strict SLOs across dozens of data centers. Design systems for multi-tenant traffic shaping, tuning both resource allocation and request handling — including smart rate limiting and