
Already know some of this? A 2-minute test skips the lessons you know.
Take the placement testA free 29-lesson course built from Sheryians AI School's 4h 47m of YouTube video. 3 days at 90 minutes a day, with study notes and a quiz in every lesson. A 2-minute placement test skips what you already know.
Taught by Sheryians AI School on YouTube. Videos play in YouTube's own player, so the creator keeps every view.
Perform EDA: check shape, dtypes, missing values, target distribution, class imbalance, outliers via boxplots, and feature correlations.
Remove duplicates, filter unrealistic ages (>100) and employment lengths, verify interest rate range, and flag rows where numbers don't add up.
Separate features (X) and target (y), define numerical/categorical column lists, and perform a stratified train-test split to maintain class ratios.
Detect class imbalance in loan_status, compute scale_pos_weight for XGBoost, and explain how class_weight in Logistic Regression balances minority class focus.
Create sklearn pipelines for numerical (impute + scale) and categorical (impute + one-hot) columns; combine them with ColumnTransformer to prevent data leakage.
Explain 5-fold stratified cross-validation, train Logistic Regression and XGBoost baseline models via cross_validate, and collect accuracy, precision, recall, F1, ROC-AUC per fold.
Build a reusable evaluate_model function that prints confusion matrix, classification report, ROC curve, and precision-recall trade-off for any estimator.
Define parameter distributions for XGBoost (n_estimators, max_depth, learning_rate, subsample, colsample_bytree, min_child_weight, gamma) and run RandomizedSearchCV with 5-fold CV.
Extract class-1 probabilities, compute precision/recall/F1 per threshold, and select the threshold (0.99) that maximizes F1 score.
Calibrate the tuned XGBoost model using CalibratedClassifierCV (sigmoid) and IsotonicRegression; compare calibration curves and Brier scores.
Install SHAP, compute global feature importance (beeswarm/bar) and local waterfall/force plots for individual predictions to see per-feature contributions.
Analyze false positives (safe people marked risky) and false negatives (risky people marked safe) on the best calibrated model to understand model struggles.
Create main.py with FastAPI, define LoanApplication Pydantic model for input validation, load calibrated model + threshold pickle, and build /predict endpoint returning risk label, probability, and threshold.
Prompt an LLM to generate HTML/CSS/JS UI, save under static/ folder, mount StaticFiles at '/' in FastAPI, and add CORS middleware for frontend-backend communication.
Create GitHub repo, push code with .gitignore, create Render web service, set build/start commands, fix Python version and requirements.txt, and verify live deployment.