Hi, I'm Yash Karle

📌 Business Problem

A telecom provider faced rising customer attrition with no systematic way to predict which customers would leave or why. Retention campaigns were untargeted, wasting budget on low-risk customers while high-risk accounts churned silently.

📂 Data

• Source: Internal CRM — 1,993 unique customer records
• Features: Demographics, service details, financial metrics, usage patterns, security features
• Target: Customer status (Churned / Stayed / Joined)

🔬 Methodology

Data Cleaning (missing values, type casting) → Feature Engineering (encoding, scaling) → XGBoost classifier with hyperparameter tuning (200 estimators, max_depth=6, lr=0.05) → Risk segmentation into Low / Medium / High tiers → SQL-based cohort analysis → Interactive Power BI dashboard.

💡 Key Insights

• 611 customers (30.7%) flagged as high churn risk — enabling proactive outreach
• Month-to-month contracts are the #1 churn driver; two-year contracts slash churn significantly
• Value Deal 5 is the strongest retention lever (18.4% feature importance)
• Customers without online security churn at materially higher rates
• Top churn reasons: competitor offers, network reliability, pricing

📈 Business Impact

• Identified ~$500K+ monthly revenue at risk from high-churn segment
• Enabled targeted retention campaigns, potentially reducing churn by 15–20%
• Automated risk scoring replaced manual quarterly reviews, saving ~40 analyst-hours/quarter
• Power BI dashboard provided leadership with real-time churn visibility for strategic planning

📌 Business Problem

An insurance provider needed a data-driven approach to understand which policyholder characteristics drive premium costs, enabling more accurate pricing and risk segmentation instead of relying on broad actuarial tables.

📂 Data

• Source: Insurance claims dataset — 2,771 records × 7 features
• Features: Age, gender, BMI, number of children, smoker status, region, charges
• Target: Insurance charges (continuous)

🔬 Methodology

Data Wrangling (replace invalid markers, impute missing values, type casting) → Correlation Analysis (identified smoker as r=0.789 with charges) → Progressive model comparison: Simple LR → Multi-feature LR → Polynomial Pipeline → Ridge Regression → Ridge + Polynomial (degree=2) with 80/20 train-test split.

💡 Key Insights

• Smoking status is the dominant cost driver (r=0.789), explaining 62% of variance alone
• Polynomial features boosted explanatory power from 75% → 85% (in-sample)
• Ridge regularisation delivered a robust R²=0.821 on unseen data
• BMI and age contribute additive, non-linear effects on premium costs

📈 Business Impact

• Enabled risk-adjusted premium pricing with 82% prediction accuracy
• Smoking-based segmentation alone could improve pricing precision by ~20%
• Model supports actuarial teams in identifying high-cost customer profiles before policy issuance
• Reduced underwriting processing time through automated risk scoring

📌 Business Problem

The sales team relied on manual spreadsheet analysis to track KPIs, leading to delayed reporting, inconsistent metrics, and no single source of truth. Leadership needed a self-refreshing dashboard for real-time performance visibility.

📂 Data

• Source: StoreSales.json — 51,291 sales records
• Features: 24 columns covering order details, customer data, location, product info, and financial metrics
• Database: SQLite with single normalised sales table (13.3 MB)

🔬 Methodology

Extract: Ingest JSON via Pandas → Load: Write to SQLite with schema auto-creation → Transform: SQL aggregate queries to compute 5 KPIs (Total Sales, Total Profit, Quantity Sold, Avg. Discount, Profit Margin) + category breakdowns and Top-10 products → Serve: Flask web app with HTML dashboard.

💡 Key Insights

• $12.6M total sales with $1.47M profit across 178K units
• Average discount of 14.29% across all transactions
• Technology category leads revenue; Office Supplies leads volume
• Identified Top-10 products contributing disproportionately to profit

📈 Business Impact

• Dashboard replaced ~8 hours/week of manual Excel reporting
• Provided leadership with real-time KPI visibility for quarterly planning
• Automated ELT pipeline ensures consistent, reproducible metrics
• Scalable architecture supports additional data sources and KPIs

📌 Business Problem

Public health agencies needed to quantify pollution trends across NYC neighbourhoods to allocate monitoring resources, prioritise interventions, and forecast future air quality under current regulatory policies.

📂 Data

• Source: NYC Dept. of Health — 17,743 records × 12 features
• Indicators: PM2.5, NO₂, O₃, boiler emissions, vehicle miles, health impacts
• Coverage: 42 UHF neighbourhoods across NYC, 2009–2022

🔬 Methodology

Date parsing → IQR-based outlier removal (Q1=8.9, Q3=26.9) → Missing value handling → Filter PM2.5 subset → Feature engineering (extract year) → 70/30 train-test split → Linear Regression → Temporal trend analysis + geographic heatmaps.

💡 Key Insights

• PM2.5 levels show a consistent declining trend from 2009–2022
• Bronx & Harlem neighbourhoods (High Bridge, Hunts Point) have the highest mean AQI values — 39.3 vs. city average of 21.6
• Seasonal patterns: Ozone spikes significantly in summer months
• Data completeness is excellent — 99.95% non-null records

📈 Business Impact

• Identified priority neighbourhoods for additional monitoring station deployment
• Forecasting model supports regulatory compliance planning and EPA reporting
• Geographic disparity analysis informs equitable public health resource allocation
• Trend projections help estimate long-term healthcare cost savings from declining pollution

📌 Business Problem

Real estate stakeholders — agents, investors, and assessors — needed a data-driven pricing model to validate property valuations, identify underpriced opportunities, and reduce reliance on subjective comparable market analysis.

📂 Data

• Source: Connecticut Real Estate Sales (data.gov) — 1,141,722 records
• Features: Town, residential type, assessed value, sales ratio, date listed
• After cleaning: 738,541 valid records (80/20 split → 590K train / 148K test)

🔬 Methodology

Date parsing → Drop rows with missing critical values → Feature engineering (Years_Since_List) → OrdinalEncoder for categorical features → XGBoost Regressor (hist method, max_depth=3, 200 estimators, lr=0.1, subsample=0.8) → Evaluation with R², MAE, RMSE.

💡 Key Insights

• Town (37.6%) and Assessed Value (33.3%) drive 71% of prediction power — location is king
• Sales Ratio (16.7%) captures market-to-assessment discrepancies
• MAE of $35,357 — usable for initial screening and bulk valuation
• Model performs strongest on mid-range properties; high-end outliers increase RMSE

📈 Business Impact

• Automated bulk property valuation for 700K+ transactions
• Supports investors in identifying undervalued properties where sale price < model prediction
• Reduces time-to-valuation from hours to seconds per property
• Assessed-vs-predicted gap analysis informs tax assessment appeals

📌 Business Problem

Before any predictive model can succeed, the underlying data must be understood. These EDAs answer foundational questions: Which neighbourhoods face the worst pollution? How well do government property assessments match actual market prices?

📂 Data

• Air Quality: ~1,000 environmental monitoring records — pollutant indicators, geographic locations, time periods
• Property Sales: ~10,000 real estate transactions — assessed values, sale amounts, sales ratios, property types

🔬 Methodology

Data Ingestion (CSV/Excel via Pandas) → Null handling & type conversion → Univariate analysis (histograms, KDE) → Bivariate analysis (scatter, box plots) → Correlation heatmaps → Statistical summarisation → Publication-ready Matplotlib/Seaborn charts.

💡 Key Insights

• Certain NYC neighbourhoods have 2× the pollution levels of city-wide averages
• Property assessments systematically undervalue properties in high-demand towns
• Box plots revealed extreme outlier sales that summary statistics missed entirely
• Seasonal air quality variation follows predictable, actionable patterns

📈 Business Impact

• EDA findings directly informed feature selection for subsequent ML models
• Property analysis revealed systemic valuation bias — actionable for government assessors
• Demonstrated a reproducible EDA framework applicable to any new dataset

📌 Business Problem

With the proliferation of AI-generated imagery, media organisations and security teams need automated tools to flag potentially manipulated facial images before they're published or used for fraud.

📂 Data

• Source: Curated dataset of real and AI-generated (deepfake) facial images
• Preprocessing: Resized to 128×128, normalised pixel values (0–1), 80/20 train-test split
• Model size: ~38 MB

🔬 Methodology

Image preprocessing (resize, normalise, RGBA→RGB) → Custom 3-layer CNN (32→64→128 filters, MaxPooling, Dropout=0.5) → Binary classification (sigmoid output) → Trained with Adam optimiser (lr=0.001) + Binary Crossentropy → Evaluated with confusion matrix & classification report → Deployed via Streamlit with real-time upload & prediction.

💡 Key Insights

• Sub-second inference enables real-time screening of uploaded images
• Confidence scoring provides probability-based risk assessment rather than binary yes/no
• Dropout regularisation (0.5) prevents overfitting on limited training data
• Research paper documents methodology for academic reproducibility

📈 Business Impact

• Deployable tool for media verification workflows and content moderation
• Streamlit interface enables non-technical users to perform deepfake audits independently
• Research contribution advances the deepfake detection knowledge base
• Architecture is extensible to video detection and ensemble methods