From Data to Model: How an AI Project Actually Runs

From Data to Model: How an AI Project Actually Runs

AI projects fail for boring reasons: an undefined question, data that does not exist, or a model deployed with nobody watching it. The modelling itself is rarely the hard part. A project that follows a clear lifecycle has a far better chance of producing something the business can use.

Define the question

Before data, decide what decision the model will support and how being right or wrong will be measured. "Improve support" is not a machine learning problem. "Route incoming tickets to the correct team" is, and it comes with a metric: the share routed correctly without human correction.

Collect and understand the data

Data is gathered from where the work already happens: tickets, orders, records, documents, logs. Then the honest questions: how much is there, how messy, how biased towards particular cases, and whether it may be used for this purpose. Projects are often reshaped at this stage, and that is a good outcome: it is much cheaper to change a plan than a production system.

Prepare and train

Splitting data properly, cleaning labels, handling missing values, choosing a baseline before a sophisticated model. A simple model that beats the current process is worth more than a complex model nobody can explain.

Evaluate honestly

Test on data the model has never seen, check the cases that matter to the business rather than the average, and decide in advance what accuracy is good enough to be useful. Comparison against how the task is done today is the only benchmark that counts.

Deploy and monitor

A model in a notebook is not a product. It needs an API, a place to run, a way to receive input and return output, and monitoring for the drift that happens when the world changes around it: new products, new language, new behaviour. Retraining is scheduled, not improvised.

How we approach it

AI and machine learning work at Black Origin IT follows discovery, planning, design, development, testing, deployment and support, with human review built in wherever an error would be costly.

Have a question a model could answer? Tell us about your project and we will tell you what the data says.