Machine learning & data science models built on your own data

A language model such as ChatGPT writes text and answers questions, and an AI agent is built on top of one to hold a conversation and take actions. Machine learning in the sense of this service answers a different question: predicting a number or classifying a record from the business's own historical data - how many units will sell next week, which lead will close, which transaction is unusual. These models are usually smaller, cheaper to run and easier to verify, because every prediction can be compared with what actually happened. Many projects combine both kinds.

It depends on the question. Sales forecasting needs a history of orders or sales by date and product, ideally two years or more so seasonality and holidays show up. Lead scoring needs past leads with a known outcome - closed or not. Document classification needs examples that were already sorted by hand. The data does not have to be tidy: an export from the CRM, the store or a spreadsheet is enough for a first review. What really blocks a project is data that was never recorded, such as a cancellation reason nobody logs. In that case the review ends with a recommendation of what to start collecting.

Not always, and the data review is where we say so. With a hundred customers and one product, a good dashboard and a few simple rules will deliver almost all of the value. Machine learning starts paying off when there are many repeated decisions over many items - hundreds of products in stock, thousands of leads a year, a steady stream of documents - and no person can look at each one. Even then we start with one question and a clear metric rather than a platform. Often the first and most useful step is a data pipeline and a dashboard, with the model coming after it.

We test it on data it has not seen. In the proof of concept we hold back the most recent period, for example the last three months, train on what came before and compare the predictions with what really happened. The same comparison is made for the method you use today, so you see a real gap rather than an accuracy percentage without context. After launch the check continues: every prediction is logged and compared with the outcome, and a periodic report shows whether accuracy is holding. When it drops, for example after a price change or a new audience, we retrain on current data.

Amendment 13 to Israel's Privacy Protection Law broadened the obligations of businesses that hold personal data, so privacy is part of the specification rather than an appendix. In practice we work on data minimisation: a sales forecasting model does not need names or phone numbers, and identifiers are replaced with codes before data leaves your systems. Data stays in an environment you control, access is logged, and working copies are deleted at the end of the project. We do not give legal advice; if your database needs a privacy adviser's review, we prepare an exact description of what is collected, where it is stored and why.

A model running inside your systems, not a file that stays with us. That usually includes the data pipeline that feeds it, the code for the model and for retraining, a connection to wherever the prediction is used - CRM, inventory, store or dashboard - and a document explaining which data goes in, what comes out and how to read the result. You also get the proof-of-concept report with the baseline comparison. The code and the model belong to you. After launch you can leave monitoring and retraining with us, or take a handover session and run it yourselves.