🧠 Stemeta Internship — ML Projects Portfolio

🏢 Projects developed during AI/ML Internship at Stemeta.ai, Islamabad Covering real-world classification problems across healthcare, finance, telecom, and HR domains — with full EDA, feature engineering, model training, evaluation, and API deployment.

---

📁 Projects Overview

#	Project	Domain	Algorithm	Key Technique
1	🔬 Breast Cancer Detection	Healthcare	Random Forest, SVM	Binary Classification
2	💳 Credit Card Fraud Detection	Finance	Random Forest, Logistic Regression	SMOTE — Imbalanced Data
3	💡 Health Indicator Analysis	Healthcare	ML Classification	EDA + Patient Profiling
4	❤️ Heart Failure Prediction	Healthcare	Logistic Regression, XGBoost	Precision-Recall Optimization
5	🧑‍💼 IBM HR Employee Attrition	HR Analytics	Decision Tree, Random Forest	Attrition Factor Analysis
6	📧 Spam Email Detection	NLP	Multinomial Naive Bayes	TF-IDF Text Vectorization
7	📱 Telco Customer Churn	Telecom	Logistic Regression, Decision Tree	Churn Prediction
8	🚢 Titanic Survival API	Classic ML	Random Forest	Flask REST API Deployment
9	⚖️ Imbalanced Data — Churn	Telecom	Multiple Models	SMOTE, Under/Over Sampling

---

🔍 Project Details

1. 🔬 Breast Cancer Detection

Supervised classification model predicting whether a tumor is benign or malignant based on diagnostic imaging features.

• Dataset: Wisconsin Breast Cancer Dataset
• Models: Random Forest, SVM, Logistic Regression
• Focus: High recall to minimize false negatives (missed cancers)

---

2. 💳 Credit Card Fraud Detection

Detecting fraudulent transactions in a highly imbalanced dataset where fraud cases are less than 0.2% of all transactions.

• Challenge: Extreme class imbalance
• Solution: SMOTE oversampling + undersampling techniques
• Models: Random Forest, Logistic Regression
• Metrics: Precision, Recall, F1, ROC-AUC

---

3. 💡 Health Indicator Analysis

Exploratory analysis of general health data to identify key disease indicators and support healthcare decision-making.

• Focus: EDA, feature correlation, patient profiling
• Techniques: Univariate & multivariate analysis, visualization

---

4. ❤️ Heart Failure Prediction

Predicting likelihood of heart failure from clinical records — where precision and recall are critical for life-saving decisions.

• Dataset: Heart Failure Clinical Records Dataset
• Models: XGBoost, Logistic Regression, Random Forest
• Focus: Optimizing recall — missing a positive case is costly

---

5. 🧑‍💼 IBM HR Analytics — Employee Attrition

Predicting which employees are likely to leave the organization and identifying the key factors driving attrition.

• Dataset: IBM HR Analytics Dataset (1,470 employees)
• Models: Decision Tree, Random Forest
• Insights: Feature importance analysis for HR decision support

---

6. 📧 Spam Email Detection

NLP-based binary text classification system to detect spam emails with high accuracy using classical NLP techniques.

• Technique: TF-IDF Vectorization
• Model: Multinomial Naive Bayes
• Libraries: NLTK, Scikit-learn
• Focus: Text preprocessing, stop word removal, stemming

---

7. 📱 Telco Customer Churn Prediction

Predicting customer churn for a telecom company to help retention teams proactively target at-risk customers.

• Dataset: IBM Telco Customer Churn Dataset
• Models: Logistic Regression, Decision Tree
• Features: Demographics, account info, service usage data

---

8. 🚢 Titanic Survival Prediction API

End-to-end ML project with REST API deployment — predicting passenger survival and serving predictions via Flask.

• Model: Random Forest Classifier
• Deployment: Flask REST API
• Input: Passenger features → Output: Survival prediction (JSON)

# Example API call
POST /predict
{
  "Pclass": 1,
  "Sex": "female",
  "Age": 28,
  "SibSp": 0,
  "Fare": 100
}
# Response: {"survived": 1, "probability": 0.92}

---

9. ⚖️ Handling Imbalanced Data — Customer Churn

Deep-dive study into resolving class imbalance in the Telco Churn dataset and measuring its impact on model performance.

• Techniques Compared:
  • SMOTE (Synthetic Minority Oversampling)
  • Random Oversampling
  • Random Undersampling
  • Combined Approach
• Finding: SMOTE + Random Forest gave best F1 score

---

🛠️ Tech Stack

Category	Tools
Language	Python 3.x
ML Library	Scikit-learn
Data Analysis	Pandas, NumPy
Visualization	Matplotlib, Seaborn
NLP	NLTK, TF-IDF
Imbalanced Data	imbalanced-learn (SMOTE)
API Deployment	Flask
Notebooks	Jupyter Notebook
Version Control	Git, GitHub

---

🚀 Getting Started

# Clone the repo
git clone https://github.com/tashfeen786/STEMETA_Intership_Projects.git
cd STEMETA_Intership_Projects

# Install dependencies
pip install -r requirements.txt

# Open any project notebook
jupyter notebook

---

🏗️ Project Structure

STEMETA_Intership_Projects/
│
├── breast_cancer/                    # Breast cancer classification
├── Credit Card Fraud Detection/      # Fraud detection + SMOTE
├── HealthIndicator/                  # Health data EDA & modeling
├── Heart_Failure_Prediction/         # Heart failure risk prediction
├── IBM HR Analytics.../              # Employee attrition analysis
├── Spam_email_Detection/             # NLP spam classifier
├── Telco_Customer_Churn/             # Customer churn prediction
├── TitanicAPI/                       # Flask API deployment
├── Handling_imbalanced_data_...ipynb # Imbalanced data study
└── requirements.txt                  # Python dependencies

---

📈 Key Learnings from Internship

• ✅ Real-world data is messy — EDA and preprocessing matter most
• ✅ Class imbalance is a critical challenge in production ML
• ✅ Model selection depends on domain — healthcare needs high recall
• ✅ Deployment bridges the gap between ML model and real product
• ✅ Feature engineering often matters more than model choice

---

👨‍💻 Author

Tashfeen Aziz — AI/ML Engineer & Python Developer

---

⭐ If you found these projects helpful, please give it a star!

Built during AI/ML Internship at Stemeta.ai, Islamabad 🇵🇰