Sr. Software Engineer, Training/Inference (C++)

At SpaceXSI we are building the high-performance inference platform that serves Grok to millions of users every day. As a Software Engineer, Training/Inference (C++), you will design and optimize large-scale model serving systems end-to-end. You will own everything from distributed infrastructure – global KV cache, continuous batching, load balancing, auto-scaling – to deep low-level optimizations, including GPU kernels, quantization, speculative decoding, and tail latency. This is a high-impact role where your work directly determines how fast and reliably users interact with Grok at massive scale.

• →Architect and implement scalable distributed infrastructure for model serving (load balancing, auto-scaling, batch scheduling, global KV cache)
• →Optimize latency and throughput of model inference under real production workloads
• →Build and operate high-concurrency serving systems with excellent tail latency and high reliability at the scale required to serve millions of users
• →Benchmark, fine-tune, and accelerate inference engines (including low-level GPU kernel work and code generation)
• →Develop custom tools to trace, replay, and fix issues across the full stack — from orchestration down to GPU kernels
• →Create robust CI/CD infrastructure for seamless endpoint deployment, image publishing, and inference engine updates
• →Accelerate research on scaling test-time compute, RL rollout, and model-hardware co-design for next-generation systems

• Bachelor’s degree in computer science, computer engineering, or science discipline
• 5+ years of experience with low-level system programming experience in C/C++ or Rust

• Experience with large-scale, high-concurrent production serving
• Experience with GPU inference engines (e.g., vLLM, SGLang, Triton, TensorRT-LLM)
• Strong background in system optimizations: batching, caching, load balancing, parallelism
• Low-level inference optimizations: GPU kernels, code generation
• Algorithmic inference optimizations: quantization, speculative decoding, distillation, low-precision numerics
• Experience with testing, benchmarking, and reliability of inference services
• Experience designing and implementing CI/CD infrastructure for inference