Models that survive contact with reality.

Applied data science · Updated September 2026

From an experiment to an operating product

My data-science work connects a model's output to the person who needs to use it. That means defining the target, the information available at prediction time and the evidence needed before a result can move beyond research. A good aggregate score is a starting point; slices, failure cases and reproducibility decide whether it is useful.

Running a live WNBA model follows the collection, evaluation and promotion cycle behind a read-only advisory product. It is not an automated trading system. When a backtest meets the real world examines comparable experiments, execution assumptions and the gaps that remain in forward-testing evidence.

Computer vision adds another kind of uncertainty: the labels themselves can be wrong. The snooker error review separates missed objects from extra detections and annotation disagreements. Reading the ice connects geometry and perception to an inspectable curling scene, while keeping automated-reference agreement distinct from human-ground-truth accuracy.

The earlier studies below cover forecasting, simulation and interactive probability tools. For the full current collection, browse quantitative research or sports intelligence.