Research
Advancing classification where it matters most
Most of the world's business and scientific data is tabular. Our research asks how to classify it accurately, efficiently and transparently — on the data organisations actually have.
Focus areas
What we study
Classification on real-world data
Real tables are messy: mixed types, missing values, skewed classes and shifting
distributions. We study methods and preprocessing choices that hold up under those
conditions rather than on cleaned benchmarks alone.
Reliability & uncertainty
A prediction is only useful if you know how far to trust it. We work on calibration and
uncertainty estimates that stay meaningful, so a "90% confident" model is right about
90% of the time.
Small data, strong results
Many high-value problems have hundreds or thousands of rows, not millions. We focus on
approaches that learn efficiently from limited examples without overfitting to them.
Evaluation & fairness
Accuracy on one split says little. We use evaluation that reflects deployment —
out-of-time checks, per-group performance and honest baselines — so results transfer.
Open questions
Questions that guide our work
- How well do current methods hold up when tables mix numbers, categories, text and missing values?
- When can a model trained on a few hundred rows be trusted — and how do we show it?
- How should confidence be reported so that people act on it correctly?
- Which evaluation practices catch the failures that matter before deployment?
Research notes
Notes & write-ups
Findings, experiments and methods, written up as we go.
Our first research notes are in preparation. Check back soon.
Working on a hard classification problem?
We are interested in problems where the data is difficult and the decision matters. Tell us about yours.