← DemosLive · login required

Credit score

An agent builds the profile and you watch the MLflow champion's pipeline run step by step, or validate and score a CSV in batch.

mlops-platform

A real credit model, step by step

It's the mlops-platform pipeline: the MLflow @champion, trained on German Credit and served by the internal scoring service. In live mode, you build a profile and watch each step run on the server. In batch mode, you validate and score a whole CSV.

1,000credit-g cases
18input variables
6live steps
250rows per CSV

…

What happens in each step

  1. 1

    Raw data

    The agent (or the form) sends a JSON with 18 fields, uncorrected in the browser: a number may arrive as text and a field may be empty. Sex and foreign-worker status aren't even asked, because they're left out of the model.

  2. 2

    Data transform

    The service casts the types, fills missing values with the credit-g median or mode and runs the 11 categoricals through the champion's own OneHotEncoder. A category rare in training falls into the "infrequent" bucket.

  3. 3

    Normalization

    The 7 numeric variables go through the champion's StandardScaler: z = (x − mean) / sd, with the training mean and sd. That puts months and reais on the same scale.

  4. 4

    Live fit

    To show the training is real, a logistic regression is trained on the spot on 750 rows and measured on 250 it didn't see (AUC). It decides nothing: it's the baseline.

  5. 5

    Inference

    The champion, the model that passed the MLflow gates, computes the log-odds; the sigmoid turns it into a probability of default and the cutoff decides between approving and sending to review.

  6. 6

    What weighed

    Each variable is swapped for its typical credit-g value and the model runs again. The difference, in percentage points, shows what pushed the risk up or down.

German Credit (OpenML credit-g, UCI) comes from a 1970s German bank, with amounts in Deutsche Marks, converted to reais (R$ 9.70 per DM) before training. Sex and foreign-worker status are left out of the model, and the performance gap between age groups is one of the promotion gates. The answer comes from the @champion in MLflow.