When I started in Data Science, my focus was primarily on improving technical skills like programming and model building. After a few years, my interest shifted towards model deployment and MLOps, leading to my transition into Machine Learning Engineering. Public speaking and presentations were always part of the job, especially when conveying results to a non-technical audience. However, things changed last year when I started working on more complex projects with potential reputational or financial risks for the hiring companies.
At this point, models required validation by a committee of both technical and non-technical reviewers before going live in production. This necessitated proper documentation, covering everything from architecture and training methodology to performance reports and experiment history. It meant that having good performance was not enough; I had to convince others, from data scientists to risk assessment specialists, that my models were not only effective but also safe.
In essence, I had to learn how to sell them.
Initiating the process of detailing my models initially served as a crucial requirement for validation, but it swiftly evolved into a routine ingrained in my approach, extending even to personal projects. Within this article, I aim to impart five valuable insights derived from my documentation experience that may assist you in crafting your own.
This is the foundation of all your documentation. When you start developping a model, you engage in a trail and error process during which you try different kind of pre-processing, model architectures, hyperparameters, and feature engineering. I highly recommend logging everything that you do, not necessarily to show everything that you found, but because you might be asked to provide an explanation of the choices you made during development.
For instance, you might have found that XGBoost models generally outperform RandomForests on your use case — if you have experiments showing that in your logs, you can easily pull them if you are asked to give an example or to show that…