Research Engineer, Model Inference & Serving - London

Key Responsibilities

• Build and operate the inference stack that serves H's multimodal agentic models
• Improve latency, throughput, and cost of model serving across the stack
• Research and implement inference techniques tailored to agent workloads
• Co-design with the Models team on training-time decisions that affect inference
• Collaborate with cross-functional teams to integrate inference into agentic AI products
• Evaluate inference, serving, and hardware platforms, and communicate findings to stakeholders
• Stay current with advancements in inference, model serving, and accelerator technology

Requirements

• Technical skills:
• Strong software engineering track record
• Proficient in Python and at least one systems language (Rust, C++, or Go)
• Hands-on experience with deep learning frameworks (PyTorch, JAX), preferably in an industry setting
• Solid distributed systems fundamentals
• Experience working in a modern cloud environment and with production ML infrastructure (Kubernetes, etc.)
• Working knowledge of modern ML, including transformers and multimodal architectures
• Research skills:
• Research engagement: an advanced degree with research output, or publications at top-tier AI or systems venues (e.g., NeurIPS, ICML, MLSys, OSDI), research internships, or substantive open-source contributions
• Soft skills:
• Excellent communication and presentation skills
• Strong collaboration and teamwork skills
• Passion for inference and AI
• Preferred qualifications:
• Startup experience
• Hands-on experience with inference frameworks (vLLM, SGLang, TensorRT-LLM)
• Writing or modifying GPU kernels (CUDA, Triton, etc.)
• Edge or on-device inference experience (llama.cpp, MLX, ONNX Runtime, etc.)
• Experience with quantization, speculative decoding, disaggregated inference or KV-cache compression
• Experience with multimodal models and/or agentic systems

Location

• Paris or London.
• This role is hybrid, and you are expected to be in the office 3 days a week on average.
• Please expect some travel between offices on a reasonable cadence (e.g., every 4-6 weeks).