Portfolios
A wider sample of shipped work, each with the metric it was judged on. For the full model cards, read the case studies.
SKU demand forecasting
Beat a seasonal naive baseline by 3.5 points of MAPE.
Warehouse capacity planning
Peak day staffing planned from a distribution, not a point estimate.
Delivery ETA prediction
Reports its own uncertainty, so dispatch knows when to distrust it.
Surface defect detection
Annotation protocol rebuilt before a single model was retrained.
Packaging label verification
Edge deployment at 22 ms per unit.
Component counting from tray images
Replaced a manual count at end of shift.
Loan document extraction
Abstains rather than invents. The abstention rate is published.
Support ticket triage
Routing beats the rules engine it replaced, on the classes that matter.
Policy document assistant
Retrieval with mandatory citation. No fine tune.
Feature store and training pipeline
Reproducible runs, versioned data, rollback in minutes.
Drift monitoring rollout
Alerted on a genuine distribution shift six weeks after launch.
Independent evaluation of a vendor model
Reproduced the vendor's claim, then evaluated on unseen data.
No work listed in this discipline yet.
Why the last entry says the claim was not replicated
Because it was not. A client asked us to evaluate a model they had been sold. We reproduced the vendor's reported metric on the vendor's own split, then evaluated on data the vendor had never seen. The performance did not hold.
We list that engagement here because a portfolio containing only successes is not a portfolio, it is an advertisement. The most valuable thing we sold that client was a number they did not want.
Ask for a reference
Tell us the shape of your problem and we will introduce you to a client who had one like it, including one we advised not to proceed.
Request a referenceSend us your data problem. We will tell you whether it is one.
A working data scientist reads every enquiry. You will get an honest read on whether your data can support the model you have in mind, before anyone quotes you a number.