Turn potential into product.

A millisecond stand-in for a one-second simulator

Design sweeps, fits and experiment planning ask the same question thousands of times. Even a fast simulator sets the pace at that point.

Built our own device simulator from the published physics, designed for thousands of runs in a row (tens to hundreds of times faster per run than the established reference, and agreeing with it to a fraction of a millivolt). Then trained a stand-in model that reproduces it across 36 device parameters, with an error budget set beforehand and met.

About 1 ms per answer instead of about 1 s, and tens of thousands of answers per second when batched. Sweeps and inverse fits run at interactive speed. Results shared with a research group; not yet in their pipeline.

Methods and software developed independently by DataDrivenLab, applied to open questions in perovskite photovoltaics. See Products.

Spectroscopy fitting that keeps up with the measurement

Fitting transient-absorption spectra meant running the same pump cycle hundreds of times until it settled. One change inside the solver cut a cycle from a quarter of a second to about a millisecond, and a reformulation removes most of the cycles altogether. The fit now keeps up with the measurement.

Methods and software developed independently by DataDrivenLab, applied to open questions in perovskite photovoltaics. See Products.

Nanoparticle size and shape from a single spectrum

Inferring the size and shape distribution of gold nanorods from one UV-Vis-NIR spectrum, by decomposing it over about 10,700 simulated optical responses. The decomposition is a non-negative least-squares fit done by gradient descent in JAX, restricted to the longitudinal plasmon region above 795 nm. The method builds on a published approach and on simulation data from other authors.

Methods and software developed independently by DataDrivenLab. See Products.

Compressing a hyperspectral camera's output on the GPU

A compression package for large measurement datasets, built for streams of gigabytes per second: hyperspectral cubes of hundreds of bands first, other imaging and spectral streams designed in. Written and tested as a standalone package, separate from the explorer above, with the question of how far the data can be compressed before it stops answering questions.

Methods and software developed independently by DataDrivenLab. See Products.

From raw camera files to an explorable dataset

An autonomous materials lab was producing hyperspectral images of hundreds of printed samples. Every look at the data started with 15–30 lines of file-parsing code.

A one-line domain language for loading a sample cube, and an interactive dashboard with false-colour RGB, wavelength sliders, a band browser and a pixel-spectrum explorer. A first explorer in a day; the version that shows the state of every sample took about a week.

An explorer that shows the state of every sample in the lab; now running as a demo there.

Deep learning from the instruments the lab already owned

A diagnostics start-up was building a deep-learning pipeline to read results from its benchtop instruments. Models had to hold across devices, bacterial strains and operators, and the validation strategy had to stand up to the FDA.

• Analysed the raw instrument data and the measurement literature; showed the existing hardware held signal an upgrade would not add.
• Designed the experiment plan that crossed devices, strains and operators so models generalise in the field.

A €100K+ equipment upgrade avoided. The instruments the lab already owned held the signal; the models drew it out.

The data-science foundation a start-up still runs on

• Designed the foundation data library and the validation strategy for the FDA submission.
• Hired the first data scientist and handed over.

Designed in two months, the validation approach and the data library became the basis of the start-up's data science and deep learning, carried it through its next funding round, and are still in use.

Real-time street geometry on the phone

Huawei Maps needed to detect street geometry from the phone camera, in real time, on the device. No cloud round-trip, no battery drain.

• Designed an energy-efficient deep network that met the accuracy target inside the on-device compute and power budget.
• For a separate, classical computer-vision model, built the team's GPU infrastructure, ML environment and CI/CD on Kubernetes.
• Led the team through research, productisation and hand-over to the Maps product.

Real-time on-device processing shipped in Huawei Maps on 100M+ smartphones; a production ML platform the team kept using.

Reconstructing outdoor scenes from sparse LiDAR

Deep probabilistic model combining PointNet, U-Net and VAE architectures. Production pipeline on Kubernetes processing terabytes of imagery a week at 98% accuracy across weather conditions and countries.

Models behind a multi-country balance sheet

Probabilistic graphical model for network risk of large lending customers, hierarchical time-series forecasting, graph neural network on a 1 TB transaction set, recommender for marketing. Deployed for €120B balance-sheet forecasting across multiple countries.

Repeatability of solenoid valves for automotive damping control

The customer needed every valve to respond the same way. I scoped a fixed-scope analysis of test-stand data: system and physics description, exploratory analysis of the variables, recommendations for data collection, and a first model, with results shared as slides. The aim was to report hysteresis and its variance per part and batch, and how sensitive those numbers are to design parameters.