Who you are

• Inference Optimization. Deep understanding of modern serving frameworks and techniques like vLLM or TRT-LLM
• Model Acceleration. Hands-on experience with quantization, distillation, caching strategies , continuous batching, paged attention, and speculative decoding
• High-Performance Systems. Proficiency in C++, CUDA, Rust, or highly optimized Python. You know how to profile code and squeeze every ounce of performance out of NVIDIA GPUs
• Distributed Systems & Scaling. Experience with Kubernetes, Ray, custom load balancing, multi-GPU/multi-node inference, and reliably handling thousands of concurrent connections
• Public work. Non-trivial systems programming projects, open-source contributions to major inference engines, or deep-dive technical write-ups
• Full-cycle ownership. You can take a model from the research team, containerize it, optimize its serving, and ensure it runs reliably in production
• Background. PhD in CS, Physics, Math, or equivalent practical experience building backend or ML systems
• You don’t need a roadmap to start walking; you’re comfortable picking a direction and building the map as you go
• You believe engineering isn't finished until it’s shipped and stable. You have a bias for impact over purely theoretical optimizations
• You don't just ship code; you obsess over the why. You’re the first to question an architecture if you think there’s a better way to solve the core latency or throughput problem
• You aren't satisfied with "the PM said so." You thrive on deep context and want to understand the fundamental logic behind every decision we make

What the job involves

• A year ago, reliably working agentic systems and sub-second multimodal inference at scale barely existed. Nobody has a decade of experience here
• So we're not screening for a resume template — we're looking for strong people from varied backgrounds who learn fast, thrive in ambiguity, and can show us what they've built, broken, and understood
• We hand you unclear problems and expect you to make them clear. We value engineers who say "I don't know yet" and then design the benchmark or prototype that finds out
• We treat performance, latency, and reliability as first-class product features, not a box to check before launch. Impact comes before everything else, though we support sharing work and open-source contributions that move the field forward. Your work should be visible. Flat structure, fast iterations, minimal process theater