GlassBox AutoML

GlassBox AutoML is a transparent machine learning library built from scratch with NumPy. The aim is to provide an end-to-end AutoML pipeline that remains readable, explainable, and easy to debug.

This is my fork of a 5-person UM6P team project. Upstream: Gyro0/GlassBox-AutoML-Agent · Package: glassbox-automl-agent on PyPI

My contribution (Mohammed Khatiri)

I was the top committer (11 of 33 commits). I built:

• the EDA module: profiler, automatic column typing, correlation and outlier detection
• the matrix helpers, the classification and regression metrics, and the consolidated test suite
• the MCP server that exposes AutoFit to LLM agents, and the IronClaw web demo
• the scikit-learn comparison benchmark below

Results against scikit-learn

Run on 29 Oct 2026 with benchmarks/sklearn_comparison.py and its default train/test split. Every GlassBox model lands within 1% of its scikit-learn equivalent.

Regression, California housing (20,640 rows, R²)

Model	GlassBox	scikit-learn
Linear regression	0.640	0.640
Decision tree	0.621	0.621
Random forest	0.691	0.686
k-nearest neighbours	0.709	0.709

Classification, stroke prediction (5,110 rows, accuracy)

Model	GlassBox	scikit-learn
Logistic regression	0.951	0.951
Decision tree	0.948	0.949
Random forest	0.951	0.951
Gaussian naive Bayes	0.571	0.571
k-nearest neighbours	0.947	0.947

Only 4.9% of the stroke rows are positive, so predicting "no stroke" every time already scores 0.951. This table shows that GlassBox reproduces scikit-learn, not that the models predict strokes well.

Project Goal

The project covers the full machine learning workflow inside the glassbox/ package:

• exploratory data analysis
• preprocessing
• models
• evaluation
• hyperparameter optimization
• agent-level AutoFit integration

Core library modules are built from scratch with NumPy only. No Scikit-Learn code belongs inside glassbox/.

Installation

From PyPI

Install the published package:

python3 -m pip install glassbox-automl-agent

PyPI project page: glassbox-automl-agent.

Manual Installation From Repository

Clone the repository, create and activate a virtual environment, then install dependencies:

git clone https://github.com/Gyro0/GlassBox-AutoML-Agent.git
cd GlassBox-AutoML-Agent
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt

If you need local development tools:

python3 -m pip install pytest jupyter

Quick Start

Run the full AutoFit pipeline on the included sample CSV:

from glassbox.agent import auto_fit

report = auto_fit(
    "data/sample.csv",
    target_column="purchased",
    task="auto",
    search="random",
    time_budget=20,
)

print(report["best_model"])
print(report["cv_score"])
print(report["eda_summary"]["overview"])

The returned report is JSON-safe and includes:

• EDA overview, numerical profile table, correlations, and outlier rows
• selected task type
• candidate model leaderboard
• best model, best parameters, and cross-validation score
• feature importances or coefficient-style importances when available

Manual Workflow Example

import numpy as np

from glassbox.preprocessing import OneHotEncoder, SimpleImputer, StandardScaler
from glassbox.models import RandomForestClassifier
from glassbox.evaluation.classification import classification_report

X_num = np.array([
    [22.0, 32000.0],
    [24.0, np.nan],
    [42.0, 76000.0],
])
X_cat = np.array([["basic"], ["basic"], ["plus"]], dtype=object)
y = np.array([0, 0, 1])

X_num = SimpleImputer(strategy="mean").fit_transform(X_num)
X_num = StandardScaler().fit_transform(X_num)
X_cat = OneHotEncoder().fit_transform(X_cat)
X = np.hstack([X_num, X_cat])

model = RandomForestClassifier(n_estimators=10, max_depth=4, random_state=42)
model.fit(X, y)
predictions = model.predict(X)

print(classification_report(y, predictions))

Model Zoo

Classification:

• LogisticRegression
• DecisionTreeClassifier
• RandomForestClassifier
• GaussianNaiveBayes
• KNearestNeighbors(task="classification")

Regression:

• LinearRegression
• DecisionTreeRegressor
• RandomForestRegressor
• KNearestNeighbors(task="regression")

Demo And Benchmarks

Launch the notebook:

jupyter notebook notebooks/demo.ipynb

Run the Scikit-Learn comparison benchmark for regression:

python benchmarks/sklearn_comparison.py --task regression --csv data/_uploaded.csv --target Delay

Run the Scikit-Learn comparison benchmark for classification:

python benchmarks/sklearn_comparison.py --task classification --csv data/classification.csv --target stroke

If --target is omitted, the script uses the last column in the CSV as the prediction target.

Scikit-Learn is used only in the benchmark script. The glassbox/ package itself remains NumPy-only.

Repository Structure

GlassBox-AutoML-Agent/
|-- .github/
|   `-- pull_request_template.md
|-- glassbox/
|   |-- agent/
|   |-- eda/
|   |-- evaluation/
|   |-- models/
|   |-- optimization/
|   |-- preprocessing/
|   `-- utils/
|-- tests/
|-- notebooks/
|-- benchmarks/
|-- data/
|-- README.md
|-- pyproject.toml
`-- requirements.txt

Testing

Run the test suite from the repository root:

python3 -m pytest -q

IronClaw / MCP Tool

The library exposes a single tool, auto_fit, through three surfaces:

• glassbox.agent.mcp_server — a FastMCP server over stdio, the IronClaw deployment target.
• glassbox.agent.mcp_tool — a JSON-in/JSON-out CLI shim for scripted/sandbox testing.
• mcp.json — reference manifest documenting the tool schema (not consumed by IronClaw directly).

Register with IronClaw

After SSHing to your IronClaw box and pip install -e .[mcp]:

ironclaw mcp add glassbox \
  --transport stdio \
  --command python \
  --arg -m --arg glassbox.agent.mcp_server

IronClaw stores the registration in ~/.ironclaw/mcp-servers.json and spawns the server over stdio whenever the agent calls the tool. Verify with ironclaw mcp list.

Run the tool directly (path mode)

python -m glassbox.agent.mcp_tool --input '{
  "csv_path": "data/sample.csv",
  "target_column": "purchased",
  "task": "auto",
  "search": "random",
  "time_budget": 20
}'

Run the tool inside a sandbox where there is no host filesystem (bytes mode):

python -c "import base64,json,sys; \
  print(json.dumps({'csv_b64': base64.b64encode(open('data/sample.csv','rb').read()).decode(), \
                    'target_column':'purchased'}))" \
  | python -m glassbox.agent.mcp_tool

The response is a single JSON object: {"ok": true, "report": {...}} on success, or {"ok": false, "error": "..."} on failure. The report includes an explanation array of short bullets the agent can repeat back to the user.

Agent private key

Never commit the IronClaw agent private key. Use one of:

1. Environment variable: export IRONCLAW_AGENT_PRIVATE_KEY=... (or place it in a local .env — already gitignored).
2. The IronClaw CLI's own keystore at ~/.ironclaw/credentials (preferred for production).

The glassbox/ package itself never reads the key; only the IronClaw runtime does, when registering the agent.