Helen Wall

The business stakeholders and users are also a big part of this process, and they interact with the outputs of the AI models in different ways. For example, speed is important and should be continually measured in recommendation algorithms. But if someone uses a ChatGPT model to get legal insights they're probably more interested in relative accuracy than speed. So the metrics used should reflect the business goals and end users of the AI model output.

Security should be one of the first steps and something that's monitored regularly throughout the process as well. The data should come from a secure source like a database and there should be security for both the model and the data it uses along the way. While setting up security often seems tedious, dealing with it earlier makes the development that comes after it much easier and smoother.

Data visualizations can often help clarify the relationship between the data and the model outputs for different algorithms. For example, we can use scatter plots to visualize groups for clustering algorithms as well as outliers for the data. We can use line charts to show anomalies in time series data (especially for moving average modeling and the impact of seasonality) and to show forecasts into the future. We can use bar or column charts to visualize the impact of different variables in a model like the PCA algorithm.

For those working on AI models in academia, there's a lot more emphasis on the theory and math behind these models than there is in industry positions. Industry roles typically focus a lot more on the business value of these AI models and the technology (like cloud infrastructures) that support them. I haven't made this transition myself, but if I did then I would take a project I already worked on within academia and apply it to a business use case. This includes quantifying its impact to a potential business problem and explaining why it brings values to an organization within industry.

Much like any model, the outputs of an NLP model are only as good as the inputs we put into it. If the initial text that we're trying to summarize is incoherent and contains a lot of confusing wording and grammatical errors, the summarization will only continue to magnify that. It's also important to check the outputs of any model as well, whether that's a number or in this case written text. I first do a passthrough to make sure the generated text represents what I actually want to say. I then do a second passthrough where I read it out loud to ensure the summarization is grammatically correct and sounds smooth and coherent!

Between getting data from its source and creating visuals, transforming the data is another key part of this process. I've found this is a step that I very rarely have the luxury of skipping along the way! Once I get the data into a table format, I'll then make sure the fields themselves are the correct data types (dates, numbers, etc.). I might need to split one field into two or more new fields, and I might need to add a new field to the table, depending on the fields that I ultimately need to create the data visualization (which is often the end product of this process)!