Research Engineer, Model Inference & Serving at H Company

You will build and operate the high-performance inference stack powering H Company’s multimodal agentic models. By optimizing latency, throughput, and cost across layers like vLLM and SGLang, you’ll ensure real-time performance for vision-action loops. You will co-design with the Models team, implementing cutting-edge research techniques to solve unique agent-specific serving constraints.

• Own the end-to-end inference infrastructure, improving the latency and throughput of the stack that serves foundational multimodal models in production.
• Research and implement specialized inference techniques tailored for agentic workloads, focusing on long contexts, tool-call latency, and vision-plus-action constraints.
• Collaborate with the Models team to influence training-time architecture choices and evaluate emerging hardware platforms to shape long-term technical direction.

• Strong software engineering background in Python plus a systems language like Rust, C++, or Go, with deep experience in distributed systems.
• Hands-on expertise with PyTorch or JAX, multimodal architectures, and inference frameworks such as vLLM, SGLang, or custom GPU kernels.
• Proven research pedigree through an advanced degree, substantive open-source contributions, or publications at top-tier venues like NeurIPS, ICML, or MLSys.

• Join an elite team of AI talent, many with DeepMind pedigree, working at the absolute frontier of computer-use agents and superintelligence.
• H Company is backed by €200M in funding from tier-1 investors including Accel, Amazon, UiPath, Eric Schmidt, and Xavier Niel.
• Translate cutting-edge research into production daily, utilizing techniques like speculative decoding and disaggregated inference to power the open-source Holo model family.