What Andrew Ng's Data-Centric AI Method Taught Our Machine Learning Agent

Andrew Ng has trained more machine learning engineers than probably anyone alive. Through Coursera, DeepLearning.AI, Stanford's CS229, and a library of free courses, he has taught the fundamentals of ML to millions of people. But the idea that shaped this blog post — and shaped how we built the dw-ml agent — is not in those curricula. It is a reframing he started pushing publicly around 2021, and it cuts against the instinct that most trained ML engineers have when a model underperforms.

The instinct is to reach for the model. Try a bigger architecture. Tune the learning rate. Add dropout. Swap the optimizer. For the first decade of deep learning, this instinct was often correct — architectures were actively improving and model choice genuinely mattered. Ng's argument is that in 2026, for most production applications, that era is over. The architecture is no longer the bottleneck. The data is.

What Is Actually Worth Learning

Ng defined his approach precisely. He called it data-centric AI: "the discipline of systematically engineering the data needed to successfully build an AI system." The word systematically is load-bearing. The problem he was diagnosing was not that practitioners ignored data — everyone knew data mattered. The problem was that data work was ad hoc, intuition-driven, and hard to reproduce. The code side of ML had git, experiment tracking, and hyperparameter search. The data side had a folder called 'cleaned_v3_final'.

His framework rests on three interlocking principles:

• •Error analysis before data collection. Sample the model's misclassifications, bucket them by category, and calculate which failure mode accounts for the most error. That category — not a random expansion of the training set — is where you invest next.
• •Targeted slice improvement over broad collection. "Collecting more data often helps, but if you try to collect more data for everything, that can be a very expensive activity." The discipline is to go narrow and deep on the highest-impact slice rather than wide and shallow across everything.
• •Consistency as the primary quality standard. Volume is not the goal; coherence is. "Having 50 thoughtfully engineered examples can be sufficient to explain to the neural network what you want it to learn." Label consistency — every annotator drawing the same boundary around the same ambiguous case — matters more than having ten thousand examples labeled in five different ways.

There is a fourth principle that underlies all three: iterativeness is not optional. "Improving the data is not a preprocessing step that you do once. It's part of the iterative process of model development." This reframes the entire ML workflow. The loop is not train → evaluate → tune model → repeat. It is train → error analysis → fix the data → retrain → compare → repeat.

How a Method Becomes a Skill

The dw-ml agent has a skill called data-centric-iteration. Its trigger conditions recognize the phrases that appear when a team is stuck in model-centric thinking: 'should we try a different architecture?', 'accuracy is stuck at 82%', 'the fraud model keeps misclassifying this category.' The skill intercepts that moment and redirects it.

The workflow the skill encodes follows Ng's loop closely. It starts with evaluate_model across slices — not a single aggregate accuracy number, but a breakdown by segment, class, or data subpopulation. Then get_feature_stats pulls examples from the worst-performing slice. The agent buckets the misclassifications manually or by detectable attribute and produces a frequency-impact table: which error category is most common, and what would a fix be worth?

Only after that analysis does the skill create an experiment — via create_experiment — and only on the targeted data fix. The model architecture is held constant. This is the heart of Ng's method: isolate the variable. If you change the data and the model at the same time, you cannot tell which intervention worked. The skill enforces this discipline by design.

After training on the improved data slice with train_model, compare_experiments tells you whether the targeted slice improved without regressing elsewhere. If it did, the data intervention is validated and logged. If it regressed other slices, the fix introduced a distribution shift and should be rolled back and re-scoped. The skill has explicit decision points for both outcomes.

One of More Than 400

Data-centric-iteration is one of more than 400 method-named skills across 19 agents in the Data Workers swarm. Each skill distills a working method from published, public-domain thinking and encodes it as a decision procedure the agent can execute reliably. The dw-ml agent alone has skills covering experiment design, feature engineering, model drift detection, canary rollouts, model explainability, and RAG architecture — each one encoding a different expert's best practice rather than a generic checklist.

The reason we build skills this way is the same reason Ng's framework works: specificity beats generality. A skill that fires on 'model not improving despite data changes' and knows to check label consistency before recommending architecture changes is more useful than a skill that says 'improve your data quality.' The method is the value.

A note on this post: This is independent commentary and homage. It distills publicly available writing and talks by Andrew Ng to illustrate a working method, and every quote is drawn from and verified against the primary sources linked above — specifically the IEEE Spectrum Q&A (https://spectrum.ieee.org/andrew-ng-data-centric-ai) and the DeepLearning.AI keynote transcript. The skill it describes is named for the method, not the person, and contains no marketing claims attributed to him. Data Workers is not affiliated with, sponsored by, or endorsed by Andrew Ng. If you are Andrew Ng and would like anything adjusted or removed, email hello@dataworkers.io and we will respond promptly.