AI Inference Engineer

In this role you will advance inference efficiency for brain-inspired AI at an early-stage research lab. You will work with cross-disciplinary teams to profile, model, and optimize cost and performance of large models across architectures and hardware. You will tackle challenges from quantisation to custom kernels, influencing core technology and energy-efficient AI at scale. This position offers impact-driven research with real-world deployments and opportunities to publish and contribute to open-source.

• Inference profiling and cost modelling across compute, memory, and bandwidth
• Develop and apply efficient inference methods (quantisation, sparsity, pruning, distillation, compression)
• Optimize inference (KV cache management, speculative decoding, batching, efficient attention) and implement GPU-focused performance improvements
• Explore non-Transformer and brain-inspired architectures and hardware-aware ML approaches
• GPU programming and performance engineering (CUDA, Triton, kernels, profiling)
• Advance on-device and edge inference and energy modelling
• Collaborate with researchers to push core technology forward

• Bachelor’s, Master’s or PhD in Maths, CS, Physics, Engineering, Neuroscience or related quantitative field
• Strong understanding of modern DL and LLM architectures and their cost drivers
• Proficient Python and PyTorch/JAX with hands-on profiling/optimisation experience on GPUs/accelerators/CPUs
• Foundations in linear algebra, probability, optimisation, numerical methods, information theory, dynamical systems or statistical learning
• Rigorous experimental mindset and ability to connect theory, experimentation and implementation
• Research publications, open-source contributions or substantial independent projects valued

• rigorous experimental mindset
• ability to move between mathematical reasoning, experimentation and implementation
• Python
• PyTorch/JAX
• model profiling and performance optimisation on GPUs/accelerators/CPUs
• Understanding of attention, KV caching and autoregressive decoding