Senior Software Engineer - Model Performance
- Compensation
- Not disclosed
- Location
- San Francisco
- Work setup
- Onsite
- Seniority
- Senior
- Employment
- Not specified
- Experience
- 2+ years
- Category
- Software engineering
- Posted
Skills
Must have
- 2+ years of experience in ML systems, inference optimization, or GPU programming
- Strong proficiency in Python and familiarity with C++
- Hands-on experience with LLM inference frameworks (vLLM, SGLang, TensorRT-LLM, or similar)
- Deep understanding of GPU architecture and experience profiling GPU workloads
- Familiarity with LLM optimization techniques (quantization, speculative decoding, continuous batching, KV cache management)
- Experience with PyTorch and understanding of how models execute on hardware
- Track record of measurably improving system performance
Nice to have
- Experience with CUDA programming
- Familiarity with serving non-LLM models (TTS, vision, embeddings)
- Experience with distributed inference and multi-GPU serving
- Contributions to open-source inference frameworks
- Experience with Docker and Kubernetes
Full description
Help us make inference blazingly fast. If you love squeezing every last drop of performance out of GPUs, diving deep into CUDA kernels, and turning optimization techniques into production systems, we'd love to meet you.
About
trains and hosts specialized language models for companies that need frontier-quality AI at a fraction of the cost. The models we train match GPT-5 accuracy but are smaller, faster, and up to 90% cheaper. Our platform handles everything end-to-end: distillation, training, evaluation, and planet-scale hosting.
We are a well-funded ten-person team of engineers who work in-person in downtown San Francisco on difficult, high-impact engineering problems. Everyone on the team has been writing code for over 10 years, and has founded and run their own software companies. We are high-agency, adaptable, and collaborative. We value creativity alongside technical prowess and humility. We work hard, and deeply enjoy the work that we do. Most of us are in the office 4 days a week in SF; hybrid works for Bay Area candidates.
About the Role
You will be responsible for making our inference stack as fast and efficient as possible. Your work spans from implementing known optimization techniques to experimenting with novel approaches, always with the goal of serving models faster and cheaper at scale.
Your north star is inference performance: latency, throughput, cost efficiency, and how quickly we can bring new model architectures into production. You'll work across the full inference stack—from CUDA kernels to serving frameworks—to find and eliminate bottlenecks. This role reports directly to the founding team. You'll have the autonomy, a large compute budget, and technical support to push the limits of what's possible in model serving.
Key Responsibilities
- Implement and productionize optimization techniques including quantization, speculative decoding, KV cache optimization, continuous batching, and LoRA serving
- Deep dive into inference frameworks (vLLM, SGLang, TensorRT-LLM) and underlying libraries to debug and improve performance
- Profile and optimize CUDA kernels and GPU utilization across our serving infrastructure
- Add support for new model architectures, ensuring they meet our performance standards before going to production
- Experiment with novel inference techniques and bring successful approaches into production
- Build tooling and benchmarks to measure and track inference performance across our fleet
- Collaborate with applied ML engineers to ensure trained models can be served efficiently
Requirements
- 2+ years of experience in ML systems, inference optimization, or GPU programming