Ankit Anand / ankitanand.ai

Chapter 1 · The Deployed Data Scientist

The Mindset Shift

A production model is not a file. It is a living data product with dependencies, users, failure modes, and an operating cost.

The first step toward reliable machine learning is changing what counts as “done.” In a notebook, success can mean a promising metric, a compelling chart, or a demo that runs from top to bottom. In production, the model is only one component in a chain that includes data acquisition, validation, features, deployment, monitoring, feedback, and ownership. A useful model that cannot be reproduced, inspected, or updated is not yet a dependable product.

This shift is not a rejection of experimentation. It is an expansion of the unit of thought. The data scientist still asks whether a model learns a meaningful signal, but also asks what happens when the input schema changes, a source arrives late, traffic doubles, or a business policy changes. Those questions turn a one-time artifact into an engineered system.

From project to product

Data products have lifecycles. They are versioned, tested, documented, observed, and retired. They need a clear owner and a defined consumer. The “accidental data guy” who inherits a pipeline at the end of an analysis is often discovering these requirements under pressure; a deployment-minded team brings them forward while the design is still cheap to change.

Reproducibility is the bridge. Pin the data assumptions, capture the environment, record the model inputs, and make the training path executable by someone else. Then define what the system promises: an input contract, a latency target, a quality threshold, and a response to failure. These are not administrative extras. They are how a team preserves trust when the original author is unavailable.

Questions to carry forward

← Return to the canonical book page