Raza Mehar's Projects
Utilized deep learning systems to classify brain MRI scans into glioma tumor, meningioma tumor, pituitary tumor, or no tumor. We addressed class imbalance using undersampling and augmented the dataset with rotation, shifting, shearing, zooming, and flipping techniques.
Analyzed employee turnover (Jan 2022 - Mar 2023) at my former organization, considering trends, departmental attrition, and tenure insights. Used predictive analytics from the 2022 Employee Engagement Survey to identify groups with flight risk. Incorporated Survival Analysis for temporal patterns, guiding decisions to improve retention.
Analyzed 157 US Energy stocks (Jan-Dec '23), identified Bullish/Bearish trends and risk categories. Used KMeans, Hierarchical, Spectral Clustering, revealing balanced returns and low volatility. Integrated data with Kafka for seamless subscriptions.
A machine translation project featuring RNN-based Seq2Seq, Transformer model, and pretrained models for translating English to Spanish and Urdu.
Analyzing Fater company's diaper market potential and enhancing revenue estimation for Naples stores: A Socio-Demographic, Territorial, and Points of Interest Perspective
This project aims to predict bank customer churn using a dataset derived from the Bank Customer Churn Prediction dataset available on Kaggle. The dataset for this competition has been generated from a deep learning model trained on the original dataset, with feature distributions being similar but not identical to the original data.
This project demonstrates image similarity search using two advanced techniques: K-Nearest Neighbors (KNN) and Approximate Nearest Neighbors (ANNOY). This project uses the Caltech 101 dataset to extract features from images using the ResNet50 model, and then performs similarity searches to identify and visualize similar images.
This project implements semantic image segmentation using two popular convolutional neural network architectures: U-Net and SegNet. Semantic image segmentation involves partitioning an image into multiple segments, each representing a different class.
Sentiment analysis on the IMDB dataset using Bag of Words models (Unigram, Bigram, Trigram, Bigram with TF-IDF) and Sequence to Sequence models (one-hot vectors, word embeddings, pretrained embeddings like GloVe, and transformers with positional embeddings).
A simple Python implementation of a neural network to solve the XOR problem using various initialization techniques and activation functions.
R-based statistical analysis of Boston Housing Data. Explored feature scales, computed descriptive stats, visualized data, and identified outliers (e.g., higher crime rates in specific areas). Examined variable relationships, calculated correlation coefficients, and presented findings via cross-classifications.
The CGI2Real_Multi-Class_Image_Classifier categorizes humans, horses, or both using transfer learning from Inception CNN. Trained on synthetic images, it can also classify real ones.
This project conducts a thorough analysis of weather time series data using diverse statistical and deep learning models. Each model was rigorously applied to the same weather time series data to assess and compare their forecasting accuracy. Detailed results and analyses are provided to delineate the strengths and weaknesses of each approach.