Why this section matters
A model sitting in a notebook helps nobody. Much of the AI work companies hire for is about putting models to work inside real products: a loan app that gives a decision in seconds, a farm advice line that answers questions, a dashboard that flags unusual transactions for a person to check. Adverts in Accra for AI engineers and AI application engineers describe exactly this, often more than inventing new models. Some of the most employable AI skills are therefore software and judgement skills, not only maths.
Two ways to put AI to work
- Use a hosted model through an API. A company runs a large model on its servers, and your app sends text and receives answers over the internet. This is the fastest way to add summarising, translation or question answering, and you pay for what you use.
- Serve your own model. You train a model, such as a credit scoring model in scikit learn, and run it yourself behind a small web API so other software can call it. You control the model and the data, but you are responsible for keeping it running and accurate.
Many real products use both: their own classic model for decisions on internal data, and a hosted language model for working with text.
What changes when a model meets real users
In a notebook, you control everything. In a product:
- Inputs are messy. People type in unexpected ways, leave fields blank and send text in several languages.