io.github.mdefrance/autocarver

AutoCarver turns raw numeric, categorical, and ordinal columns into optimal, drift-robust, human-readable bins in a few lines of code. Stop losing model performance to suboptimal manual binning — and stop discovering overfit bins in production monitoring.

• Provably optimal — exhaustive search: for a fixed min_freq, max_n_mod and metric (Tschuprow's T by default, or Cramér's V), no other admissible bin combination scores higher. It checked them all so you don't have to.
• Robust by construction — every candidate grouping is vetoed unless it holds on a held-out dev set (and optional CV folds), at fit time rather than in monitoring.
• Define → carve → model — declare your Features, fit a carver, transform: the whole feature set is carved in one supervised pass, not one notebook per feature. One carver per target type — BinaryCarver, MulticlassCarver, OrdinalCarver, ContinuousCarver (regression) — all with the identical API.
• AI-assisted — a local MCP server lets your LLM assistant qualify and carve columns through tool calls, fully on your machine.

On the Titanic quick start, Fare collapses from 72 pre-carving modalities to 2 bins while its association with survival rises: Tschuprow's T 0.18 raw → 0.29 carved.

Built for credit scoring, fraud detection, and risk modeling.

🆕 What's New

📊 Cross-validated robustness. fit now accepts a cv argument for extra held-out robustness views on top of (or instead of) a dev set: carver.fit(X, y, cv=5). Accepts an int, any scikit-learn splitter, or explicit index pairs, resolved via sklearn.model_selection.check_cv — folds veto over-fit combinations but never reorder them (ranks stay anchored to the full train set). See Cross-validation folds.

🤖 LLM & MCP integration. AutoCarver now ships a local Model Context Protocol server: point an MCP-aware assistant (VS Code Copilot, Claude Desktop, Cursor, …) at a data file and let it qualify the columns and carve them against your target through tool calls. The server runs fully on your machine — your dataset is never sent to AutoCarver or any external service (only your own LLM provider sees what the assistant shares). Carving quality depends on the LLM, so have a human confirm the feature definitions before production use. See the LLM & MCP guide.

pip install "autocarver[mcp]"

Once configured, just ask your assistant:

Qualify the columns in titanic.csv and carve them against Survived.

The assistant infers feature types, proposes a carving, and returns the summary table — no code written by hand.

{
  "servers": {
    "autocarver": {
      "command": "python",
      "args": ["-m", "AutoCarver.mcp"]
    }
  }
}

{
  "servers": {
    "autocarver": {
      "command": "uv",
      "args": ["run", "python", "-m", "AutoCarver.mcp"]
    }
  }
}

Install

pip install autocarver

Quick Start

You already have a DataFrame and a target — that's the first box ticked before you start:

• Load data
• Split train / dev
• Declare features by type
• Fit the carver, validated on the dev set
• Inspect the carved bins
• Persist

The rest is the snippet below — binary classification on the Titanic dataset:

from pathlib import Path

import pandas as pd
from sklearn.model_selection import train_test_split

from AutoCarver import BinaryCarver, Features

# 1. Load data
url = "https://web.stanford.edu/class/archive/cs/cs109/cs109.1166/stuff/titanic.csv"
data = pd.read_csv(url)
target = "Survived"

# 2. Train / dev split, stratified on the target
train, dev = train_test_split(data, test_size=0.33, random_state=42, stratify=data[target])

# 3. Declare features by type
features = Features(
    categoricals=["Sex"],
    numericals=["Age", "Fare", "Siblings/Spouses Aboard", "Parents/Children Aboard"],
    ordinals={"Pclass": ["1", "2", "3"]},
)

# 4. Fit the carver (dev set drives the robustness checks)
carver = BinaryCarver(features=features)
train_processed = carver.fit_transform(train, train[target], X_dev=dev, y_dev=dev[target])
dev_processed = carver.transform(dev)

# 5. Inspect the carved buckets, target rate, and association
carver.summary

# 6. Persist for later use
carver.save(Path("titanic_carver.json"))

# 7. Load the carver back in
carver = BinaryCarver.load(Path("titanic_carver.json"))
dev_processed = carver.transform(dev)

min_freq and max_n_mod are the only two knobs that matter to start with — the defaults (0.02 / 5) reflect common scoring practice, and every behavioral toggle lives in one ProcessingConfig object. Scan, adjust, move on.

For multiclass classification use MulticlassCarver (one binning per feature, against the full K-class target) — or OneVsRestCarver for a separate binning per class; for ordinal targets use OrdinalCarver; for regression use ContinuousCarver — the API is identical. To pre-select features by target association and inter-feature redundancy, pipe the carved output through ClassificationSelector or RegressionSelector with features=carver.features (the carver carves its own copy, so your original features stay raw).

What you get

Two questions worth answering before your next model review: can you defend every bin boundary of your current model to a stakeholder — and can you show each one holds on data it has never seen? AutoCarver makes both a one-liner:

• No performance left on the table — exhaustive search over admissible bin combinations maximizes Tschuprow's T (default) or Cramér's V: for fixed min_freq, max_n_mod and metric, no other combination scores higher, so you never wonder whether a better grouping existed.

• Stop silent overfitting before production — bins that only exist in your training sample degrade quietly under drift. Every candidate combination is validated on a dev set (and optional CV folds): any whose target rates flip or whose buckets fall below min_freq is rejected at fit time, not discovered in monitoring.

• First-class ordinal features — OrdinalDiscretizer enforces your declared modality order, so under-represented levels are merged with their nearest neighbour instead of being collapsed by frequency.

• You are the final auditor — carver.features.summary and carver.features.history expose the bin definitions, per-bin target rate / frequency, and the full carving trace; disagree with a boundary and you can override it, and transform applies your fix like any carved bin:

  feature = carver.features("Siblings/Spouses Aboard")  # any carved feature; labels are [0, 1, 2]
  feature.group([1], 2)  # merge two bins you consider equivalent

• Interpretable buckets — human-readable boundaries you can audit, document, and ship to a scorecard.

• Dimensionality reduction — groups under-represented modalities and caps bins per feature (max_n_mod), which is especially useful before one-hot encoding.

• Feature pre-selection — ClassificationSelector / RegressionSelector rank features by target association and filter on inter-feature correlation.

Titanic Age, one BinaryCarver.fit call: 84 raw values collapse to 3 buckets with a monotonic survival rate.

How does it compare?

All three libraries are sklearn-Pipeline compatible; AutoCarver adds JSON round-trip persistence (carver.save("...json")) and feature pre-selection helpers (ClassificationSelector, RegressionSelector). The full feature matrix, side-by-side runnable snippets, and a "when to pick which" guide live on the comparison page.

Documentation

Full reference, tutorials, and end-to-end notebook examples on ReadTheDocs.