top of page


Face Recognition Using Siamese Network

The Face Recognition Dataset is a collection derived from the Labeled Faces in the Wild (LFW) Dataset, comprising face photographs primarily of famous individuals for face detection and recognition model development. It offers RGB images of 250 x 250 pixels, organized into 1680 directories, each representing a different celebrity with 2 to 50 images per individual. Additionally, the dataset includes an "Extracted Faces" folder, featuring Haar-Cascade-processed 128 x 128 pixel faces for enhanced classifier performance.

Challenges arise from its large number of classes (1678) with limited data per class. To address these challenges, two approaches are explored: Few-Shot Learning, allowing recognition of new classes with minimal labeled examples, and Siamese Models, which learn similarity functions for comparing faces under varying conditions. The most suitable model for facial recognition will be chosen based on accuracy and generalization, using pretrained models within the Siamese Network architecture. For more information, visit the dataset's official website: Proper citation of the LFW dataset and related works is encouraged when using this dataset in research or projects.

  • LinkedIn
  • 1_Xa26ULIpZTTP1hN_ReUXVQ
  • GitHub

Custom OCR 

This project focuses on developing an Optical Character Recognition (OCR) system using Seq2Seq models with the aim of finding the best model for OCR through hyperparameter optimization using surrogate functions. Key features optimized include architecture, learning rates, convolution layers, linear layer nodes, growth relationships in convolution layers, memory unit layers, hidden size in memory units, batch size, optimizer, and Seq2Seq models.

The Seq2Seq models are employed for OCR development using the Connectionist Temporal Classification Loss (CTC Loss), which helps align sequences for tasks like audio-to-text. Surrogate functions are used to approximate the best hyperparameters for the model by exploring a search space of hyperparameters and minimizing the loss.

The dataset for training and evaluation can be customized, but a dataset of captcha images is provided for download. The dataset follows specific naming conventions, and the code is provided to create test, train, and validation data folders.

To save the best hyperparameter-trained model, the project provides code for training and saving the model, taking into account the optimized hyperparameters.

Additionally, CTC decoding techniques are mentioned for interpreting the OCR model's outputs. This project aims to build an efficient OCR system with optimized hyperparameters for various OCR tasks.

For more details and code, refer to the project repository.


Handpose Actuation Using Computer Vision

This project focuses on improving hand-pose estimation using depth images. It aims to enhance accuracy and latency optimization in comparison to existing work by Liuhao Ge. Benchmarking results indicate improved performance. Depth images from the NYU dataset, obtained from Microsoft Kinect V2, are converted into a volumetric representation. The project utilizes techniques such as Axis Aligned Bounding Boxes (AABBs), Occupancy Grids, and Truncated Signed Distance Fields (TSDFs) to process and analyze the depth data. The provided instructions outline the process of cloning research work, creating TSDFs, and training models. The project references prior work in real-time 3D hand pose estimation using convolutional neural networks.


Sentiment Analysis

This sentiment analysis project utilizes a Kaggle dataset to analyze and predict tweet sentiments as positive or negative. The dataset is sourced from Kaggle and addresses class imbalance. Data augmentation techniques such as synonym replacement and back translation are employed to augment the dataset. Preprocessing steps involve stop word removal, lemmatization, and lowercasing. The dataset is split into training and testing sets, and a Count Vectorizer with a max feature limit is used for feature engineering. The model of choice is Logistic Regression, yielding results detailed in the project. To get started, clone the repository, download the dataset, install necessary libraries, run preprocessing and feature engineering scripts, and train and evaluate the model. The project is open source, and contributions are welcome.


Breast Cancer Detection

This project is focused on breast cancer detection using neural networks. It employs a dataset with 32 columns, including an ID, diagnosis (malignant or benign), and 30 real-valued features derived from digitalized breast mass images. These features describe various characteristics of cell nuclei present in the images, such as radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, and fractal dimension.

The neural network model, implemented using Keras, consists of multiple hidden layers with ReLU activation functions, a dropout layer to prevent overfitting, and an output layer for binary classification (benign or malignant). It's trained using back-propagation and optimized with stochastic gradient descent.

To use the model, you can find the code in the provided Jupyter notebook file. The project requires Python libraries like Keras, TensorFlow, Pandas, NumPy, and Scikit-learn. The dataset used was created by Dr. William H. Wolberg at the University of Wisconsin, Clinical Sciences Center in Madison, WI, USA.

Overall, this project aims to develop a neural network-based breast cancer detection system using a dataset of image-derived features.


Document chat LLM

This project aims to achieve specific objectives in the field of Natural Language Processing ( LLM ) and chat applications. It involves the development of a chat application capable of interacting with both PDF documents and webpages.

Key components of the project include:

  1. Extracting text from PDFs and webpages.

  2. Segmenting text into manageable chunks.

  3. Generating embeddings for these text chunks.

  4. Building a conversation chain with memory within a Flask application session.

  5. Designing a user interface (UI) with HTML and CSS, integrated with the Flask backend.

The application allows users to choose between Hugging Face or OpenAI embeddings, upload multiple PDFs, read data from webpages, and delete previous chat history sessions to avoid feeding them into future queries.

To run the application, Python 3.8.16 is required. Users should navigate to the 'flask_application' directory and run 'python'. The project also includes a 'requirements.txt' file for installing necessary library packages. Additionally, users need to create a '.env' file and populate it with API keys/tokens for the chat application to function effectively.


Disease Detection 

The Disease Detection Project is a machine learning-based system designed to predict the presence or absence of various diseases. It uses a dataset containing health-related features and provides a Flask application for making disease predictions using a trained model. Users can input their health data through a web form and receive predictions. The project is open to contributions, licensed under MIT, and encourages collaboration to enhance disease detection capabilities.


Container Detection and Recognition

During my tenure as a private contractor with Netarus LLC, located in Norfolk, USA, I had the privilege of contributing as a Machine Learning Engineer to the development of a cutting-edge computer vision product. This innovative solution was meticulously crafted to address the critical task of tracking containers by utilizing Optical Character Recognition (OCR) and localization techniques. Under strict adherence to confidentiality agreements through the signing of a Non-Disclosure Agreement (NDA), I am unable to divulge the specific details or underlying codebase of the project.

Our collaborative efforts yielded remarkable results, culminating in the achievement of a remarkable 99% accuracy rate for the Machine Learning algorithms employed in real-time container tracking. This accomplishment not only signified a substantial advancement in precision and recognition capabilities but also significantly elevated the overall performance of the system.


Recommender System Development

During my tenure as a private contractor at Healthbird, I undertook the role of a Machine Learning Engineer, where I spearheaded the development of a robust recommender system for the health insurance platform. This system revolutionized the customer experience by seamlessly incorporating critical factors from user profiles, including age and income, and employing real-time statistical analysis to offer tailored insurance policy recommendations.

Collaboration was pivotal to the success of this project, as I worked closely with licensed insurance agents to create synthetic data, ensuring the system's accuracy and relevance within the complex landscape of the United States health insurance sector. This synergy between machine learning and insurance expertise allowed us to provide customers with informed choices, eliminating the need for intermediary agents.

To maintain the system's effectiveness, I integrated continuous live data feeds from the website, enabling the recommender system to evolve dynamically over time, adapting to changing customer needs and industry trends. My responsibilities extended beyond development, as I also took charge of deployment, monitoring, and ongoing application refinement, seamlessly integrating with cross-functional engineering teams to ensure a cohesive and high-performing product.

This endeavor reflects my dedication to leveraging advanced technology to enhance user experiences, streamline processes, and contribute to the evolving landscape of the health insurance sector in the United States.

The recommender syste can be used at this link:

bottom of page