Phase 1
Discovery
We pin down the decision the model will inform, the cost of being wrong, and the data realistically available to learn from.
We build prediction systems against problems with a clear cost of being wrong. Demand that has to be staffed against. Claims that have to be paid or denied. Customers about to leave next quarter. The work begins by asking what a one-percent better answer is worth — and ends with a model your team can defend in a regulator's room.
We are openly anti-hype about technique. A well-tuned gradient boost will outrun a neural network on most tabular problems, and we will tell you so. Where deep models do earn their keep, we ship them with eval harnesses, drift monitors, and the calibration curves that show when they should not be trusted.
Each one delivered with a baseline, an eval set, and the monitoring to catch the day it starts lying.
Hierarchical forecasts with prediction intervals, holiday and promotion effects, and reconciliation across SKU, store, and region.
Survival models and uplift estimators that tell you not just who will leave — but who will respond to a retention offer.
Two-tower retrieval, learning-to-rank, and bandits for the catalog problems where the answer changes every minute.
Difference-in-differences, synthetic controls, and DAG-aware estimators for the questions a correlation cannot answer.
Sequential tests, CUPED variance reduction, and an experiment registry that stops the same test from being run twice.
Drift, calibration, and segment performance tracked in production. Alerts you wired into the systems your on-call already watches.
A rhythm built around honest baselines — so we know early whether a model can actually move the number.
We pin down the decision the model will inform, the cost of being wrong, and the data realistically available to learn from.
A simple, defensible model first. Often it is most of the value — and it sets the bar everything fancier has to clear.
Feature engineering, candidate models, hyperparameter search — all tracked, all reproducible, none of it on a single laptop.
Out-of-time validation, segment audits, calibration plots, and a written argument for whether to ship or to walk away.
Send us last year's predictions and the actuals. A free read-out will tell you whether the model is broken — or the brief was.