← Back to labs
Case study · ML engineering

AutoML Lite

An automated ML pipeline for regression tasks. Upload data, train models, compare metrics, and predict — all without writing code.

automl-lite.app
3regression models
Autofeature detection
4evaluation metrics
R²best model picker
The problem

Machine learning for non-ML engineers

Most ML tools require Python knowledge, data science experience, and complex setup. AutoML Lite lets anyone upload structured data and get trained models with professional metrics — bridging the gap between raw data and actionable insights.

Architecture

Five phases, three models

Phase 1

Data ingestion & feature detection

Accepts JSON arrays via API and auto-detects numeric vs categorical columns. Handles mixed data types without manual configuration.

  • Auto-detect numeric and categorical columns
  • Handles missing values and high-cardinality strings
  • StandardScaler for numeric, OneHotEncoder for categorical
  • Automatic column filtering for usable features
Phase 2

Multi-model training

Trains three regression models in parallel: LinearRegression, Ridge (regularized), and RandomForest. Automatically tries XGBoost if available.

  • LinearRegression — simple baseline, interpretable
  • Ridge — prevents overfitting with L2 regularization
  • RandomForest — non-linear ensemble, handles interactions
  • XGBoost (optional) — gradient boosting for higher accuracy
Phase 3

Metrics & comparison

Evaluates every model on a held-out test set (80/20 split) and reports R², MAE, RMSE, and normalized MAE% for interpretability.

  • R² — how much variance the model explains (0-100%)
  • MAE — average prediction error in original units
  • RMSE — penalizes larger errors more heavily
  • MAE% — error relative to target range for intuition
Phase 4

Feature importance

Extracts which input columns most influence predictions using RandomForest feature_importances_, helping users understand what drives their data.

  • Top 10 features ranked by importance
  • Visual bar chart representation
  • Contextualizes model behavior for non-experts
  • Enables data-driven decision making
Phase 5

Prediction service

Once trained, users can input new feature values and get instant predictions from the best-performing model.

  • POST /api/automl/predict — one or multiple rows
  • Uses the best pipeline automatically
  • Returns predictions in original units
  • Models cached in memory for fast inference
Stack

Tools used

Pythonscikit-learnpandasFastAPIReactDocker