01 — Business Context & Objective

Sales managers and business executives require real-time visibility into overall company performance across multiple dimensions (Region, Product Category, Segment, and Team). The core business objective was to eliminate manual spreadsheet reporting and build an interactive decision-support system to monitor revenue, evaluate profitability, and diagnose sudden shifts in order velocity.

02 — Data Architecture & DAX Modeling

The underlying data model connects normalized transaction records with product catalogs and customer demographic entities. Multi-page drill-downs and cross-filtering required establishing strict relational cardinalities and authoring robust DAX calculations:

03 — Core Analytical Discoveries

• Fixed Margin Structure: Profit margin remained steady at exactly 30.00% across regions and products, revealing a uniform markup pricing model.
• High Customer Retention: 376 out of 500 total customers were repeat buyers (75.2% retention rate), indicating organic product-market fit.
• Category Volume Driver: Accessories significantly outperformed Electronics in total sales volume (₹3.29 Cr vs ₹2.35 Cr), establishing accessories as the company's primary cash-flow engine.
• Revenue Anomaly & Volatility: Month-over-month sales growth fluctuated drastically from +99% (April) to -60% (August), diagnosing severe order volatility rather than organic linear growth.

01 — Context & Business Problem

Rainfall prediction in India is non-linear and heavily influenced by complex interactions between temperature, pressure, wind velocity, and elevation. The objective of this project was to clean multi-state meteorological sensor feeds, uncover regional seasonal distributions, and train regression models to predict rainfall levels.

02 — Data Cleaning & Imputation Strategy

Sensor telemetry data contained over 200,000+ missing values across critical features: air_pressure (304,664 missing), wind_speed (274,444 missing), and rainfall (257,554 missing). Dropping rows would cause devastating data loss. I applied Median Imputation because meteorological variables contain severe geographic outliers, making the median far more robust than mean replacement.

03 — Analytical Honesty: Evaluating Model Overfitting

Following feature engineering and train-test splitting (80/20 split), regression models were fitted on historical readings. Evaluating the models surfaced a critical discrepancy:

Root Cause Diagnosis: Weather data exhibits strong temporal autocorrelation. A standard random train/test split caused temporal data leakage during training, while geographic climate variance between distinct states created massive distribution shifts. In production, this must be solved via TimeSeriesSplit and region-segmented models.

01 — The Reporting Challenge

Election tallies are traditionally published as massive, disjointed tables across 42 political parties and 140 legislative seats. Journalists, campaign strategists, and researchers require an interactive, politically neutral tool to compare coalition performance (UDF, LDF, NDA) and analyze localized voting behaviors.

02 — Data Schema & Hierarchy Modeling

Constructed a relational star schema normalizing districts, constituencies, party alliances, and candidate tallies. Authored custom DAX measures to dynamically calculate seat shares, winning margins, and voter turnout percentages across multi-level drill-through pages (State Overview → District Dashboards like Palakkad, Ernakulam, and Thiruvananthapuram).

03 — Key Findings

• Alliance Dynamics: UDF secured the highest aggregate state-wide vote count, while LDF maintained strong runner-up density in traditional industrial districts.
• Postal Vote Impact: Postal ballots represented a minor fraction of the total vote pool and had minimal impact on flipping constituency margins; final outcomes were almost entirely determined by EVM tallies.
• Constituency Competitiveness: Several central Kerala seats recorded victory margins of under 1.5%, visualizing fierce multi-party competition.

01 — Problem Definition & Clinical Inputs

Early identification of high-risk individuals enables preventive interventions. This project engineered a supervised classification pipeline evaluating 8 patient risk parameters:

• Demographics & Physiology: Age, Gender, Body Mass Index (BMI).
• Lifestyle Factors: Smoking Habit, Physical Activity level, Alcohol Intake.
• Clinical History: Genetic Risk Score, Family History of Cancer.

02 — Resolving Class Imbalance with SMOTE

Clinical risk datasets naturally suffer from extreme class imbalances (far fewer positive high-risk diagnoses than negative low-risk cases). Training naive classifiers resulted in deceptive 90%+ baseline accuracy that failed entirely on minority high-risk recalls.

I implemented SMOTE (Synthetic Minority Over-sampling Technique) to synthesize realistic minority examples in feature space, standardizing continuous inputs with StandardScaler to eliminate scale biases across BMI and Age.

03 — Model Selection & Deployment

Tested multiple algorithms via stratified K-Fold cross-validation: Logistic Regression, KNN, Naive Bayes, and Random Forest. Random Forest delivered optimal discrimination performance across all evaluated metrics:

• Accuracy: 95.0%
• AUC-ROC Score: 0.936
• F1-Score: 0.928

The optimized model pipeline was serialized via Pickle and deployed as a real-time interactive Streamlit web application with immediate inference capabilities.