An end-to-end FinTech ML platform that predicts EMI eligibility (classification) and the maximum safe monthly EMI (regression) for loan applicants — with full MLflow experiment tracking and a production-style Streamlit app.

---

Table of Contents

1. The problem
2. What this delivers
3. Architecture
4. Dataset
5. Feature engineering
6. EDA highlights
7. Modeling & results
8. MLflow tracking & model registry
9. The Streamlit app
10. CRUD / Data Management
11. Quickstart
12. Deploying to Streamlit Cloud
13. Project structure
14. Tech stack

---

1. The problem

People frequently take on EMIs — for phones, appliances, vehicles, education — without a clear picture of what they can actually afford, leading to missed payments and financial stress. On the lender's side, underwriting is often manual, slow, and inconsistent across loan officers.

EMIPredict AI answers two questions instantly for any applicant:

Question	ML task	Answer
Should this EMI be approved?	Classification	Eligible / High_Risk / Not_Eligible
What's the most they can safely pay per month?	Regression	max_monthly_emi (₹)

Domain: FinTech & Banking · Business impact: faster underwriting, risk-based pricing instead of blanket declines, standardized eligibility criteria across 5 lending scenarios.

---

2. What this delivers

Requirement	Delivered as
Dual ML problem solving (classification + regression)	src/train_classification.py, src/train_regression.py
Real-time risk assessment on 400,000 records	src/generate_dataset.py (synthetic, realistic, reproducible)
Advanced feature engineering (22+ variables)	src/feature_engineering.py
≥3 models per task, compared & justified	4 models each — see §7
MLflow experiment tracking + model registry	sqlite:///mlflow.db backend, 10 tracked runs + 2 registered models
Streamlit Cloud–ready multi-page app	app/ (6 pages)
Complete CRUD for financial data management	app/pages/5_Data_Management.py (SQLite-backed)

---

3. Architecture

flowchart TD
    A["📄 Raw Dataset<br/>400,000 records · 5 EMI scenarios"] --> B["🧹 Preprocessing<br/>quality checks · cleaning · train/val/test split"]
    B --> C["🛠️ Feature Engineering<br/>14 derived ratios & risk scores"]
    C --> D["📊 EDA<br/>10 figures + insights report"]
    C --> E["🤖 Model Training<br/>4 classifiers + 4 regressors"]
    E --> F["📈 MLflow Tracking<br/>params · metrics · artifacts"]
    F --> G["🏆 Model Selection<br/>best of each task"]
    G --> H["📦 Model Registry<br/>emi_eligibility_classifier<br/>emi_max_amount_regressor"]
    H --> I["🖥️ Streamlit App<br/>6 pages · real-time predictions · CRUD"]
    I --> J["☁️ Streamlit Cloud<br/>public deployment"]

    style A fill:#1565C0,color:#fff
    style F fill:#0194E2,color:#fff
    style H fill:#2E7D32,color:#fff
    style I fill:#FF4B4B,color:#fff
    style J fill:#37474F,color:#fff

Loading

Both training pipelines and the live app import the same src/feature_engineering.py module, so a prediction made in the UI goes through identical transformation logic to what the models were trained on — no train/serve skew.

---

4. Dataset

No dataset file was supplied with the project brief, so src/generate_dataset.py synthesizes 400,000 realistic financial profiles rather than using arbitrary random data. It builds an underlying financial-capacity simulation — disposable income, an affordability ratio driven by credit score, job stability, dependents and savings, plus noise — so the relationships between features and targets are realistic and learnable, not fabricated.

EMI Scenario	Records	Amount Range	Tenure Range
E-commerce Shopping EMI	80,000	₹10K – ₹200K	3–24 months
Home Appliances EMI	80,000	₹20K – ₹300K	6–36 months
Vehicle EMI	80,000	₹80K – ₹1,500K	12–84 months
Personal Loan EMI	80,000	₹50K – ₹1,000K	12–60 months
Education EMI	80,000	₹50K – ₹500K	6–48 months

22+ input variables across personal demographics, employment & income, housing & family, monthly financial obligations, credit history, and loan application details.

Realistic data-quality issues — missing values, duplicate rows — are injected intentionally so the preprocessing pipeline has real problems to clean (see reports/data_quality_report.json for the before/after audit).

---

5. Feature engineering

Beyond the raw fields, src/feature_engineering.py derives 14 features that mirror real underwriting logic:

• debt_to_income_ratio, expense_to_income_ratio, affordability_ratio
• requested_emi — computed from amount/tenure via the standard EMI formula, using a per-scenario interest-rate table
• requested_burden_ratio — requested EMI ÷ disposable income (one of the strongest predictors)
• employment_stability_score, emergency_fund_months, income_per_dependent, credit_score_norm, and more

---

6. EDA highlights

• ~51% of applicants are Not_Eligible, ~32% High_Risk, ~17% Eligible under the simulated risk rules — mirroring real retail-lending books, where most raw applications need pricing adjustment rather than a straight approval.
• Vehicle and Personal Loan EMIs (large ticket, long tenure) show the highest decline rates; E-commerce Shopping EMI (small ticket, short tenure) the highest approval rate.
• Disposable income, salary, and the engineered affordability ratio dominate the regression signal; debt-to-income and expense-to-income ratios are the strongest negative drivers.

Full report with all 10 figures + business recommendations: reports/eda_report.md

---

7. Modeling & results

4 models were trained per task (above the required minimum of 3), with identical ColumnTransformer preprocessing (one-hot + scaling) inside an sklearn Pipeline.

Classification — EMI Eligibility

Model	Accuracy	Precision (macro)	Recall (macro)	F1 (macro)	ROC-AUC (OvR)
Logistic Regression	0.8442	0.8356	0.8224	0.8279	0.9425
Random Forest	0.9584	0.9563	0.9498	0.9530	0.9732
XGBoost ✅ selected	0.9606	0.9600	0.9508	0.9552	0.9733
Decision Tree	0.9557	0.9522	0.9474	0.9498	0.9682

Held-out test set (XGBoost): 96.15% accuracy · F1 0.956 · ROC-AUC 0.974 — comfortably above the project's 90% accuracy target.

Regression — Maximum Monthly EMI

Model	RMSE (₹)	MAE (₹)	R²	MAPE
Linear Regression	2,477	1,471	0.9635	56.8%
Random Forest	1,008	563	0.9939	3.9%
XGBoost ✅ selected	978	557	0.9943	4.6%
Decision Tree	1,152	639	0.9921	4.4%

Held-out test set (XGBoost): RMSE ₹972 · R² 0.994 · MAPE 4.5% — well under the project's 2,000 RMSE target.

Why XGBoost won: eligibility and affordability are governed by thresholds and interactions (e.g. "burden ratio > 0.95 and credit score < 600 → high risk"), not straight-line relationships — which is why the linear baselines lag well behind the tree-based models.

Full metrics: reports/model_comparison_classification.md · reports/model_comparison_regression.md

---

8. MLflow tracking & model registry

Every training run — hyperparameters, all evaluation metrics, confusion matrices / prediction-vs-actual plots, feature importance, and the model artifact itself — is logged to a local MLflow tracking server (sqlite:///mlflow.db). The best model per task is registered to the MLflow Model Registry:

mlflow ui --backend-store-uri sqlite:///mlflow.db

Registered model	Source experiment	Selected on
emi_eligibility_classifier	EMI_Eligibility_Classification	Validation accuracy
emi_max_amount_regressor	EMI_MaxAmount_Regression	Validation RMSE

The app's Model Performance page queries this database live to display run history alongside the comparison tables above.

---

9. The Streamlit app

6 pages, all real-time and backed by the actual trained model pipelines:

Page	What it does
🏠 Home	Overview, live model-availability status
📊 EDA Dashboard	Interactive, filterable Plotly charts over the dataset
✅ Predict Eligibility	Real-time classification with class-probability chart
💰 Predict Max EMI	Real-time regression with an affordability gauge
📈 Model Performance	Full model comparison + live MLflow run browser
🗂️ Data Management	Full CRUD over stored loan applications
ℹ️ About	Project overview, architecture, business use cases

Run it locally:

streamlit run app/Home.py

---

10. CRUD / Data Management

The Data Management page implements full CRUD over a local SQLite store (data/app_data.db, auto-created):

• Create — add a new applicant record; predictions run automatically.
• Read — search, filter, and export all stored applications to CSV.
• Update — edit a record, optionally re-running predictions on the new data.
• Delete — remove a single record or clear the store entirely.

---

11. Quickstart (local)

# 1. Install dependencies
pip install -r requirements.txt

# 2. Generate the 400K-record dataset (~6 seconds)
python src/generate_dataset.py

# 3. Clean, validate, engineer features, and split train/val/test
python src/preprocessing.py

# 4. Run EDA (writes reports/eda_report.md + reports/figures/*.png)
python src/eda.py

# 5. Train & MLflow-track classification models (4 models)
python src/train_classification.py

# 6. Train & MLflow-track regression models (4 models)
python src/train_regression.py

# 7. Launch the app
streamlit run app/Home.py

---

12. Deploying to Streamlit Cloud

1. Push this repository to GitHub — trained models/*.pkl files and reports/ are already committed, so the app works immediately after deploy with no retraining required on the cloud instance.
2. On share.streamlit.io, create a new app pointing at this repo with main file path: app/Home.py.
3. Streamlit Cloud installs requirements.txt automatically and deploys.
4. (Optional) To retrain from scratch on a fresh clone, run the pipeline scripts from §11 before starting the app.

---

13. Project structure

EMI prediction LABMENTIX/
├── app/
│   ├── Home.py                      # Landing page
│   ├── utils.py                     # Model loading, prediction, CRUD (SQLite)
│   └── pages/
│       ├── 1_EDA_Dashboard.py       # Interactive EDA (Plotly)
│       ├── 2_Predict_Eligibility.py # Real-time classification
│       ├── 3_Predict_Max_EMI.py     # Real-time regression
│       ├── 4_Model_Performance.py   # Model comparison + MLflow runs
│       ├── 5_Data_Management.py     # CRUD for loan applications
│       └── 6_About.py               # Project overview
├── src/
│   ├── generate_dataset.py          # Synthesizes the 400K-record dataset
│   ├── preprocessing.py             # Cleaning, quality checks, train/val/test split
│   ├── feature_engineering.py       # Shared feature logic (training + app)
│   ├── eda.py                       # Figures + markdown EDA report
│   ├── model_utils.py               # Shared pipeline/MLflow/plotting helpers
│   ├── train_classification.py      # 4 classifiers + MLflow tracking + selection
│   └── train_regression.py          # 4 regressors + MLflow tracking + selection
├── data/
│   ├── raw/emi_dataset_sample.csv   # 5,000-row sample (committed, for the EDA page)
│   ├── raw/                         # Full 400K dataset (generated, gitignored)
│   └── processed/                   # train/val/test splits (generated, gitignored)
├── models/                          # Trained pipelines + metadata (committed)
├── reports/                         # EDA report, figures, model comparison reports
├── assets/                          # README graphics
├── requirements.txt
├── .streamlit/config.toml
└── README.md

---

14. Tech stack

Python · pandas · NumPy · scikit-learn · XGBoost · MLflow · Streamlit · Plotly · Matplotlib/Seaborn · SQLite

_{EMIPredict AI · FinTech & Banking Capstone Project}