Projects Portfolio

Conducted Exploratory Data Analysis (EDA) on 10,000 customer records. Resolved class imbalance using SMOTE on training data. Programmed and optimized Random Forest, XGBoost, and Logistic Regression models using Scikit-learn.

Random Forest Classifier emerged as the best model with an F1-score of 0.69 and an AUC-ROC of 0.91. The primary churn drivers identified were days since last interaction, monthly bill charges, and account tenure.

Generated a synthetic dataset of 400 images of green/red shapes with anomalies. Split data 80/20 into train/validation sets. Designed a 3-block Conv2D network with max-pooling, dense classification layers, and dropout regularization.

Achieved 100% precision, recall, and overall accuracy on the validation set, demonstrating a robust pipeline for basic visual quality inspection.

Constructed a user-item matrix of 5,000 ratings from 200 users. Utilized Cosine Similarity to calculate item-to-item correlation matrices and predicted missing user ratings based on similar movie evaluations.

Successfully generated top-10 personalized movie suggestions with predicted rating scores for any active user ID.

Analyzed 3 years of daily sales data. Conducted Augmented Dickey-Fuller (ADF) tests (p-value: 0.83) confirming non-stationarity. Decomposed trend, seasonality, and residual noise. Fitted a Holt-Winters Exponential Smoothing model.

Created a robust 28-day sales forecast accounting for weekly seasonality cycles. Achieved a Mean Absolute Percentage Error (MAPE) of 25.56%.

Preprocessed 1,000 customer reviews using tokenization, stop-word removal, and POS-tagged lemmatization (NLTK). Extracted unigram/bigram features using TF-IDF. Compared Naive Bayes, SVM, and Logistic Regression.

All three classifiers achieved 100% test accuracy. Verified that the performance is due to highly distinct linguistic templates in the synthetic dataset.

Integrated open-source LLMs (Llama and Gemma) into Python environments. Designed local inference optimization pipelines and bypassed paid cloud APIs to minimize latency.

Created an extremely responsive, 100% offline assistant executing natural language commands and complex tasks directly on personal hardware.

Programmed rendering and simulation logic in Pygame. Designed a SQLite database structure to persist NPC interactions, relationships, and history, enabling LLMs to fetch context in real-time.

NPCs demonstrate dynamic, evolving behaviors and decision-making instead of predefined script patterns, remembering past interactions.

Utilized yfinance and ccxt to fetch high-precision historical data. Trained machine learning models on pattern predictions and programmed custom entry/exit logic and backtesting environments.

Achieved accurate historical backtests and automated execution routines driven by ML signals, laying a foundation for dynamic risk management.

Programmed Playwright browser automation to handle Google Maps' complex DOM rendering. Extracted URLs, phones, reviews, and rating parameters into structured database/CSV outputs.

Drastically cut B2B lead auditing times from hours to minutes, establishing a rapid data acquisition pipeline.