This competition was organized by the HASOC (Hate Speech and Offensive Content) committee on Kaggle. The goal of this competition was to identify hate speech and offensive content (HASOC) from memes in Bangla, Hindi, Gujarati, and Bodo. Separate training and testing datasets were provided for each language. I worked on the Bangla meme dataset. A meme consists of an image and text. In the training dataset, there was an image folder and an annotated CSV file. Each meme was labeled across four categories: Sentiment (Positive, Negative, Neutral), Sarcasm (Sarcastic, Non-sarcastic), Vulgarity (Vulgar, Non-vulgar), and Abuse (Abusive, Non-abusive). This was a multimodal data fusion task requiring the integration of image and text in the model. For the image branch, I experimented with ResNet50 and ResNet152, while for the text branch, I used BanglaBERT, XLM-RoBERTa-large, and BERT-base-uncased. During training and validation, I achieved near-perfect results. For the official test set on Kaggle, the ranking metric was the average macro F1 score. I ranked 7th out of 17 teams (23 participants) with a score of 0.5763, while the top team achieved 0.62755. All code, results, and reports are available in my GitHub repository. From this task, I learned how to concatenate multimodal features and how to train a model using a multihead classifier layer. I also learned that one category could work as a cue for predicting other categories in multimodal data.

Packages: Resnet50, Restnet152, BanglaBERT, XLM-RoBERTa-large, Transformer, Tensorflow, Numpy, Pandas and Keras.

See More Details: GitHub, Report

In this project, I investigated the effect of text augmentation on model training. Text augmentation is more difficult than image augmentation. The basic units of text are words. Words have syntactic and semantic meaning that depends on the context. There are many text augmentation techniques, such as paraphrasing, rephrasing, synonym replacement and back translation. I used LLM-based semantically similar augmented data to train the model. The overall accuracy improved from 87.77% to 95.71%, nearly a 10% increase. AUC increased from 0.94 to 0.99. This experiment proves that augmentation can increase a model's generalization capacity and help classify unseen data more robustly.

Packages: Transformers, Dataset, Numpy, Pandas, Matplotlib, Sklearn, Os, Collections, Evaluate

See More Details: GitHub, Report

This project was part of the Natural Language Processing course (MSc). In English, there are many words that have multiple meanings. Meanings vary based on the context. In this project, I used Wikipedia and WordNet to disambiguate the meaning of an ambiguous word. For similarity measurement, I used Lesk, BERT, Word2Vec, GloVe, and FastText. See my GitHub README.md for more details. 

See More Details: GitHub

Generative AI could be used for identity theft and harmful activities. To prevent this, we can use another AI to detect deepfake voices and prevent crimes. In this project, I trained and fine-tuned LSTM, CRNN, and FCNN models to detect AI-generated voices. I achieved 98%, 99%, and 98% accuracy on LSTM, CRNN, and FCNN, respectively. See my GitHub README.md for more details. 

See More Details: GitHub

This project is part of the Deep Learning course (MSc). In this project, we fine-tuned pretrained models—VGG16, ResNet, EfficientNet, and DenseNet121—using a retinopathy dataset to detect retinopathy. To further improve accuracy, we used attention mechanisms and ensemble learning. We used Grad-CAM to visualize the parts of the image that the models used for training. See my project report on GitHub.

See More Details: GitHub

See More Details: GitHub

See More Details: GitHub

See More Details: GitHub

See More Details: GitHub

This project is part of the Big Data course (MSc). One year’s worth of news coverage data was collected from the GDELT (Global Database of Events, Language, and Tone) database. The collected data was refined and normalized to perform exploratory data analysis (EDA). I found a significant gap between the number of news articles mentioning Palestine and those mentioning Israel. The average sentiment toward Israel and Palestine varies greatly across different regions, such as the Middle East, Europe, and North America. The full report is on GitHub.

Tools: PySpark, Pandas, Matplotlib, Socket, GeoIP, PyCountry, TLDExtract, Plotly, Geopandas, Folium, Prophet, Cartopy.

See More Details: GitHub, Report, Presentation

Harnessing Apache Spark and advanced visualization techniques, I transformed 245K+ weather records spanning three decades into actionable insights. This project showcases my expertise in big data analytics, data cleaning, and dynamic storytelling to drive strategic decision-making. 

Tools: Apache Spark, Pandas, Matplotlib, Seaborn, Google Colab.

See More Details: GitHub

It is a wonder how a simple equation can create a complex and beautiful pattern.

See More Details: GitHub

This C++ program converts an image into an artistic sketch. See the code on GitHub.

See More Details: GitHub