What is mrpravin?

mrpravin is a Python library that automates the entire ML pipeline for tabular data — from raw CSV to trained, production-ready model — in as few as 3 lines of code.

import mrpravin as mr

df    = mr.pravinDA("data.csv")                  # clean, encode, ready
model = mr.pravinDS(df, target="loan_status")    # AutoML → best model
model.summary()                                  # full model card

No manual preprocessing. No encoder fitting. No scaler setup. No model selection loop.

Install

# Core (pandas, numpy, scikit-learn)
pip install mrpravin

# Full — adds XGBoost, LightGBM, encoding detection
pip install "mrpravin[full]"

The 3 Functions

Quick Start

Data Analyst Mode

import mrpravin as mr

# Works with CSV, Excel, JSON, or a DataFrame
df = mr.pravinDA("data.csv")

print(df.head())   # fully cleaned, encoded, human-readable
print(df.shape)

What happens automatically:

• Duplicate rows removed
• Missing values filled (median for numeric, mode for categorical)
• Outliers winsorized
• Boolean columns → 0 / 1
• Categorical columns → one-hot encoded
• High cardinality columns → frequency encoded
• Datetime columns → year / month / day / dayofweek features
• ID columns → dropped

Data Scientist Mode — AutoML

import mrpravin as mr

df    = mr.pravinDA("data.csv")
model = mr.pravinDS(df, target="price")

model.summary()

What happens automatically:

• Train / test split with zero data leakage
• Runs: LinearRegression / LogisticRegression, RandomForest, GradientBoosting (+ XGBoost, LightGBM if installed)
• Cross-validated hyperparameter tuning
• Picks the best model
• Evaluates on held-out test set
• Returns a pravinML object ready for production

Production Inference — pravinML

import pandas as pd

# Predict on new raw data
new_data = pd.DataFrame({
    "feature_1": [8, 3, 6],
    "feature_2": [95, 55, 75],
    "category":  ["Yes", "No", "Yes"],
})

predictions = model.predict(new_data)       # auto-cleans internally
probabilities = model.predict_proba(new_data)  # classification only

# Validate schema before predicting
report = model.validate(new_data)
print(report.summary())

# Feature importance
for feature, pct in model.explain().items():
    print(f"  {feature}: {pct:.1f}%")

# Benchmark inference speed
bench = model.benchmark(new_data, n_runs=100)
print(f"p50: {bench['p50_ms']} ms | throughput: {bench['throughput_rows_per_sec']} rows/s")

# Save and load
model.save("model.pkl")
model = mr.pravinML.load("model.pkl")

Real Results

Both achieved with identical 3-line code.

Configuration

from mrpravin import MrPravinConfig

cfg = MrPravinConfig(
    random_seed=42,
    cv_folds=5,
    n_iter_search=20,        # hyperparameter search iterations
    use_xgboost=True,
    use_lightgbm=True,
    outlier_method="iqr",    # "iqr" | "zscore" | "mad"
    verbose=True,
)

model = mr.pravinDS("data.csv", target="label", cfg=cfg)

Save and reuse config:

cfg.to_json("my_config.json")
cfg = MrPravinConfig.from_json("my_config.json")

Architecture

mrpravin/
├── mrpravin/
│   ├── __init__.py          ← public API
│   ├── config.py            ← MrPravinConfig
│   ├── pipeline.py          ← pravinDA() and pravinDS()
│   ├── ml.py                ← pravinML (inference layer)
│   ├── core/
│   │   ├── loader.py        ← CSV / Excel / JSON loading
│   │   ├── profiler.py      ← column type detection
│   │   ├── cleaner.py       ← dedup, imputation, outliers
│   │   ├── encoder.py       ← OHE, frequency, boolean encoding
│   │   ├── scaler.py        ← StandardScaler / RobustScaler
│   │   └── report.py        ← report builder + JSON/HTML export
│   └── automl/
│       ├── model_selector.py ← problem detection + candidates
│       ├── tuner.py          ← RandomizedSearchCV
│       └── evaluator.py      ← metrics + feature importance
└── tests/
    └── test_mrpravin.py     ← 37 tests

Full pipeline flow

CSV / Excel / JSON / DataFrame
        ↓
   pravinDA()
        ├── load
        ├── detect column types  (7 types)
        ├── clean                (dedup, impute, outliers, text)
        ├── encode               (OHE / frequency / boolean / datetime)
        └── returns DataFrame    ← human-readable, ML-ready
        ↓
   pravinDS()
        ├── dedup full dataset before split  (prevents X/y misalignment)
        ├── train / test split               (no leakage)
        ├── clean + encode + scale           (fit on train only)
        ├── model selection + CV tuning
        ├── evaluate on test set
        └── returns pravinML object
        ↓
   pravinML.predict()
        ├── validate schema
        ├── clean + encode + scale           (transform only, no re-fit)
        └── predict

Zero data leakage by design. Encoders and scalers are always fit on training data only, then applied via .transform() at inference time.

pravinML — Full API Reference

model.predict(X)              # predict labels / values
model.predict_proba(X)        # predict class probabilities
model.validate(X)             # schema + drift check → ValidationReport
model.evaluate(X, y)          # metrics on any labelled dataset
model.explain(top_n=20)       # feature importance as % contribution
model.summary()               # full model card printout
model.benchmark(X, n_runs=100)# inference latency p50/p95/p99
model.save("model.pkl")       # persist with integrity checksum
mr.pravinML.load("model.pkl") # load with checksum verification
model.metrics                 # raw metrics dict
model.feature_names           # training feature list
model.problem_type            # 'regression' | 'binary_classification' | ...
model.schema                  # InputSchema — raw feature ranges
model.model_name              # winning algorithm name

Supported Formats

Requirements

• Python ≥ 3.9
• pandas ≥ 1.5
• numpy ≥ 1.23
• scikit-learn ≥ 1.3
• scipy ≥ 1.10
• openpyxl ≥ 3.1 (Excel support)

Optional:

• xgboost ≥ 1.7
• lightgbm ≥ 4.0
• chardet ≥ 5.0 (non-UTF-8 CSV encoding detection)

Run Tests

cd mrpravin
pip install -e ".[dev]"
pytest tests/ -v

37 tests covering profiler, cleaner, encoder, scaler, pravinDA, pravinDS, pravinML (predict, validate, evaluate, explain, benchmark, save/load).

Roadmap

• Phase 1 — pravinDA · pravinDS · pravinML
• Phase 2 — pravinAI — static pipeline compiler and anti-pattern detector
• Phase 3 — pravinDL · pravinNLP — deep learning and NLP extensions

Author

Pravin MR — Data Engineer & ML Systems Builder, Chennai, India

• Website: mrpravin000.vercel.app
• Email: mrpravin000@gmail.com
• LinkedIn: linkedin.com/in/mr-pravin
• GitHub: github.com/mr-pravin

License

MIT © 2026 Pravin MR — see LICENSE for full text.