The right tool for a narrow job
Large language models get the headlines, but plenty of valuable problems are better solved by a dedicated model trained for one purpose: spotting a defect in a product photo, routing a document to the right department, predicting which invoices will be paid late, flagging a suspicious transaction. These models are small enough to run cheaply at high volume, consistent enough to audit, and often more accurate than a general model prompted to do the same thing.
Framing comes before modelling
Many ML projects fail before any code is written, because the question is vague. “Predict churn” becomes useful only when it says which customers, how far ahead, and what action the prediction will trigger. We spend the first days framing the target, the decision it supports and the metric that matters to the business — which is rarely raw accuracy.
Honest validation
The most common ML mistake is a model that looks great in testing and fails in the real world, usually because of data leakage or a test set that doesn’t resemble production. We split data the way the model will be used (often by time), check for features that “cheat”, and compare every model with a simple baseline. If a gradient-boosted tree beats a neural network, we ship the tree.
Built to run
A model is only useful in production. We package models as versioned services or batch jobs with input validation, latency budgets and logging, and we set up monitoring for data drift and performance decay. Each model ships with a model card describing what it was trained on, where it performs well and where it shouldn’t be trusted.