Deploying and monitoring responsibly
Shipping the model is the middle of the story, not the end A model sitting in a notebook helps no one. Deployment means putting it where real decisions happen — behind an API, inside a nightly batch job that scores customers, or embedded in a product feature. But going live is where the real responsibility […]
Drift: the model was right last year
A deployed model quietly decays Software either works or throws an error. A model does neither: it keeps returning confident answers while getting steadily worse, and nothing in the system raises its hand. This is drift, and it is the single most common way a successful ML project stops being one. Two kinds, and they […]
Retraining without breaking things
The new model is better on average. That is not enough Retraining feels routine — fresh data in, better numbers out, ship it. The risk is that “better on average” hides “worse on the cases that matter”, and the people affected by those cases are precisely the ones who will notice. Compare on a fixed […]
Bias is in the data before it is in the model
A model trained on past decisions reproduces past decisions Machine learning finds patterns in history. Where that history contains human decisions, it contains their biases too — and the model will learn those as faithfully as anything else, then apply them at a scale and speed no human could. Removing the field does not remove […]
Explaining a model to the people it affects
Three audiences, three explanations “Explainability” is treated as one requirement. It is three, and an explanation aimed at the wrong audience satisfies nobody. The person affected by the decision They need one thing: what would have to be different for a different outcome. “Your application was declined; with six months more trading history it would […]
Practical ML: From Idea to Model
Turn a fuzzy business goal into a concrete ML task, prepare honest data, beat a baseline, and ship a model you can monitor and trust.