Building intelligence, not noise – A Data-Centric Approach to AI

By Yuri Domingues

There is a very common belief about artificial intelligence that goes against everything I learned over the last few months working with data: that what matters is the model.

That the secret is choosing the right LLM, the newest architecture, the most sophisticated fine-tuning.

It is not.

What research has already shown

Andrew Ng, co-founder of Google Brain, former chief scientist at Baidu and professor at Stanford, has been saying this publicly for a few years. In his words: “If 80% of our work is data preparation, then ensuring data quality is the most important work of a machine learning team.”

That is not an opinion. It is a conclusion from someone who built some of the largest AI systems in the world.

A Landing AI experiment makes this even clearer. A company wanted to use computer vision to detect defects in steel plates. The baseline accuracy was 76.2%. An AI team spent two months trying to improve the model. The result: zero improvement. Ng’s team then tried a different approach, focused entirely on the quality of the training data. In two weeks, accuracy went up to 93.1%. Almost 17 percentage points. Without changing a single line of the model.

MIT formalized this approach in a full course: Introduction to Data-Centric AI. The main argument is simple: in the real world, data is not fixed. You can and should engineer your data just as much as you engineer your model.

What I saw happen in practice

Working on the development of WSB.one, a market intelligence platform for the offshore sector, the theory became reality very quickly.

The Brazilian maritime market has a lot of public data: chartering records at ANTAQ, regulatory decisions, vessel movements, corporate structures. All published. Almost nothing connected.

Building scraping pipelines with Python and Django to collect this information was challenging. Many sources were hard to access, each portal with its own behavior, its own structure, its own obstacles. But the real work was not collecting.

It was in turning raw data into something with real quality.

Standardizing vessels names that appear in different forms across different databases. Reconciling corporate entities that are split across records. Identifying when the same information, from different sources, tells contradictory stories. That is the work no model solves on its own, and it determines whether the final result is intelligence or just well-formatted noise.

What this means for those using AI today

If you are implementing AI in your company and the results are not coming, the answer is probably not in changing the model. It is in the data you are feeding it.

A few questions worth asking before any change in architecture:

Is the data consistently formatted, or does each source follow its own logic? Are there inconsistencies between sources that the model is trying and failing to reconcile? Is the data you have representative of the real problem, or is it just the easiest data to collect?

The Data-Centric AI approach does not ask you to ignore the model. It asks you to stop treating data as a detail to be fixed later.

The model is a tool. Data is the raw material. And no tool, no matter how sophisticated, turns bad raw material into a good result.