Sven
Olaf
I am a PhD student researcher and Research Software Engineer at Imperial College London, specialising in foundational machine learning and applied mathematics for scientific computing, working on sequential neural surrogates, state-space models, and simulation-based inference. I also work as a full-stack Machine Learning Research Software Engineer supporting the WHO national malaria control programmes. Founder & CEO of HiddenState. Previously at Cambridge and DFKI. Work spanning NeurIPS, Nature and The Lancet.
I make expensive computations fast and tractable. I build neural surrogates that replace costly simulations with learned emulators, design state-space architectures for efficient language and time-series modelling, and optimise LLM training through multi-fidelity Bayesian methods.
Learning to emulate computationally expensive simulations using deep temporal models. Achieving 125,000x speedup at 99.8% accuracy for epidemiological agent-based models used in WHO pandemic preparedness.
Efficient state-space architectures (Mamba-2) in JAX for causal language modelling and time-series forecasting. Contributions merged into Google's jax-ml/bonsai model zoo with up to 28x inference speedup through state-space caching.
Developing methods for likelihood-free inference in complex stochastic systems. Publications in Nature and The Lancet. NeurIPS 2025 workshop on tokenised flow matching for hierarchical SBI.
Making ML models run faster through compiler-level optimisation. Replacing custom CUDA kernels with portable XLA compilation, achieving hardware-agnostic deployment across CPUs, GPUs, and TPUs from a single codebase.
Sample-efficient multi-fidelity optimisation for expensive black-box functions. Applied to LLM data mixture discovery, jointly optimising over model scale and training duration via learning curve extrapolation.
Frameworks: JAX, PyTorch, Flax, CUDA.
Focus: State-space models, temporal emulation, Bayesian inference, DuckDB analytics.
Recognition: SAGE Award from UK Chief Scientific Officers.
Research goal:
make simulations learnable,
make inference tractable,
make science faster.
Always open to interesting problems and people.
