Rucha's Portfolio

Research Work - Master’s Thesis

My research is centered on developing novel blueprint workflow for capturing a medium sized building from acquisition using a drone to Level of Detail 3 (LOD3) standard that doesn’t require high level expertise in building modelling.
This workflow incorporates both traditional and deep learning-based computer vision techniques and models.
My work extends to generating a multi-modal UAV dataset of heritage buildings in Montreal involving data collection and its subsequent post-processing. It is soon to be open-sourced, aiming to facilitate further research.
The developed pipeline has been tested and evaluated on the collected data, playing a crucial role in the conservation and digital twinning of heritage structures, aiming to contribute towards fostering a sustainable digital future.

hr_analytics

Performed data pre-processing and feature engineering techniques on employee data, identifying key factors influencing their job decisions.
Conducted training and evaluation of multiple supervised machine learning algorithms for predicting employee retention.
Achieved a peak accuracy of 94.78% using Support Vector Machine (SVM) algorithm.

pneumonia_fewshot

This project involved using a limited and arbitrary sample from the PneumoniaMNIST dataset.
Assessed if combinations of augmentation techniques previously implemented increase the network’s performance by data preprocessing and augmentation, integration of attention modules into a CNN model.
Applied transfer learning by finetuning different models achieving 3.45% accuracy improvement over benchmarking using AlexNet.

face_mask

Generated a new dataset of facial images by collecting, pre-processing and performing data augmentation.
Trained and fine-tuned a CNN for facial mask detection and type classification.
Eliminated bias and improved model performance using k-fold cross-validation.

book_rec

Developed a recommender using goodbooks dataset of 10k records to suggest books of interest and predict user ratings.
Used Apache PySpark for Content-based and Collaborative Filtering approaches, using one-hot encoding and matrix factorization for latent feature identification.